Brett Adcock: Humanoid Run on Neural Net, Autonomous Manufacturing, $50T Market #229
Peter DiamandisDave BlundinBrett Adcock
Adcock’s core claim is that general-purpose autonomy—not manufacturing volume—unlocks a labor market the hosts size at roughly $50 trillion. Figure could finance and build 100,000 robots, he says, but that means little if they require teleoperation or replay fixed motions; “If you don’t solve that, none of this matters.” Conversely, he believes a truly general humanoid could attract demand for a billion units immediately.
Helix 2 is Figure’s claimed architectural break: the remaining 109,000 lines of C++ are gone, leaving an end-to-end neural stack for more than 40 degrees of freedom. Its learned System Zero controller coordinates full-body movement while System 1 integrates cameras, fingertip touch and palm cameras; onboard inference drives motor torques hundreds of times per second. The resulting room-scale kitchen behavior—including using a hip and foot—was something Adcock said, “You could never code.”
Figure’s prospective moat compounds through fleet data rather than task-specific software libraries. One neural network handles logistics, dishes and other work, with Adcock reporting positive transfer as diverse data is added: “Once one robot learns how to do a task, every robot in the fleet knows it.” Figure designed Helix 2 around the pre-training set, Figure 03 around Helix 2, and is bringing 3,000 B200s online for pre-training.
The strongest operating evidence offered was one error across 67 hours of continuous package work over multiple robots. The robots reportedly worked at human speed, found and positioned barcodes and even patted packages flat for the scanner; a six-month BMW deployment also ran every workday. Adcock’s real benchmark, however, is autonomous work lasting days in an unseen location—not karate, backflips or videos with “a guy in Tennessee driving it.”
Manufacturing is being scaled in parallel, with robots being targeted for Figure’s own production lines during 2026. The current facility can support four lines at roughly 12,000 units each, just under 50,000 annually, while a near-term target is one robot every 30 minutes. Longer term, Adcock discussed $10,000-$20,000 robots and hopes that within 24 months “all the robots will build all the robots,” although one billion units at $20,000 would still require $20 trillion of working capital.
The home roadmap is aggressive but explicitly staged rather than a mass-market promise. Adcock expects that by the end of 2026 Figure might place a robot in an unseen home for fairly long-horizon work, measure interventions per hour, day or week, and begin limited user deployments the following year; “I don’t want to ship slop.” He also said Figure is not yet ready for fully autonomous operation freely around his children or to hold his newborn, making that personal threshold the readiness test.
Adcock expects “far less than 10” global humanoid winners and views China collectively as Figure’s only serious competitive threat today. He nevertheless argues that closed-loop autonomy remains scarce worldwide and that every major technology company will enter because “you have no choice.” Figure intends to keep its model tied to its own vertically integrated hardware, declining to license it on safety grounds and describing safe deployment as a “fiduciary duty to our civilization.”
1. Figure went from functional prototype to robot-filled campus in 18 months
Diamandis and Blundin toured roughly 300,000 square feet, with another 400,000 under development, and estimated they saw at least 100 complete Figure 03 robots plus many more hands, heads and partial assemblies. Diamandis disclosed that his venture fund invested in two earlier Figure rounds.
Figure 01 was deliberately inelegant: an approximately 130-to-140-pound CNC-aluminum machine designed and made to walk in under a year. Adcock’s priority was to “unlock the AI and controls team,” learn actuators, batteries, wiring, structures and sensors, and provide hardware for the first bimanual neural policy.
Figure 02 internalized the wiring, added cameras, compute and roughly twice the battery capacity; Figure 03 then dropped to about 135 pounds while retaining speed and torque, carrying approximately 20 kilograms. It added soft coverings, fingertip touch, palm cameras, a passive toe and fewer pinch points.
2. The Keurig task killed the coded-robot roadmap
The decisive experiment was not a backflip but Figure 01 making coffee. A bimanual neural network picked up a K-Cup, opened the machine, inserted the pod and ran a sequence lasting several minutes—the first time Figure saw neural control work across a meaningful humanoid task.
That demonstration answered two early questions together: could Figure build a capable, affordable electric humanoid, and could it avoid “coding your way out of this problem”? Adcock’s conclusion two years ago was categorical: “We have to just go all in on neural nets. The whole stack needs to be neural nets.”
Figure previously maintained several hundred thousand handwritten C++ lines—expensive to test, difficult to deploy reliably and incapable of representing every contact-rich behavior. Helix 1 removed most of them but retained a coded lower-body controller; Helix 2 removed the final 109,000 lines.
Blundin’s distinction was that learned behavior produces useful surprises as well as mistakes. The kitchen robot used its hip to close something and its foot to lift a dishwasher door, actions the team had not scripted; Adcock’s verdict was, “You could never code this.”
3. Helix 2 closes the loop from perception to motor torque
Helix 2’s System Zero, or S0, is a full-body reinforcement-learned controller. Learned locomotion exists elsewhere, Adcock conceded, including in staged martial-arts behavior, but he had not seen it integrated with learned perception and manipulation across an entire moving humanoid.
The accompanying System 1 work integrates head and rear cameras, downward-looking torso cameras, fingertip tactile signals and Figure 03’s palm cameras. Those extra viewpoints matter when the head camera is occluded—for example, while a hand reaches into a cabinet or removes pills from a cartridge.
Inference runs fully onboard, turns sensor observations into motor torques and updates control a few hundred times per second. The robot can coordinate eyes, hands, feet, legs, pelvis and torso without waiting for a remote server, then recover from errors and replan while still handling an object.
The architectural refactor took most of a year. Figure moved from strong stationary tabletop manipulation to room-scale autonomy: walking through a kitchen, opening storage, selecting objects and putting them away. Adcock’s next spatial milestone is graduating from “finish the whole room” to the full house.
4. Figure 03 was designed as a body for Helix
Adcock described a reversed design hierarchy: Figure 03 was built for Helix, Helix 2 was built for the pre-training data, and sensors, operating system, middleware, firmware, thermals and embedded compute were chosen accordingly. “How do we give Helix a body?” became the hardware program’s governing question.
The dimensionality explains why generic hardware and text models are insufficient. With more than 40 degrees of freedom and motors that can rotate through 360 degrees, Adcock characterized the pose space as 360 to the power of 40—“more states of the humanoid than atoms in the universe.”
Figure 03’s actuators reportedly retain three-to-five times the speed headroom shown in current demos, but Adcock resisted the assumption that faster is automatically better. More arms, extreme running or maximum actuator speed add cost, mass and danger without necessarily increasing conveyor throughput; Blundin noted that a plate becomes hazardous at 5X speed.
5. Language supplies semantics, but embodiment demands its own physics
Adcock corrected the simplified story that Figure merely left OpenAI. OpenAI and Microsoft co-led Figure’s Series B, and the parties explored next-generation humanoid models, but Adcock said Figure’s internal team “ran circles around them” for much of a year and eventually saw no reason to train outsiders on its embedded-model work.
Language and vision-language models still matter: their weights encode objects, semantics and common-sense relationships, helping a robot understand that an item is a water bottle or interpret a verbal request. Adcock called that grounding “super critical” inside Helix rather than dismissing LLMs outright.
Blundin’s challenge was that an LLM may seem to know how to play soccer or grasp a bottle while lacking physical competence. Adcock agreed: it does not know the required elbow angle, fingertip pressure, pelvis motion or contact dynamics. “This is not an LLM. The LLM knows none of this.”
A zero-shot experiment at Adcock’s new Hark lab made the gap concrete. A multimodal model received a digital joystick and was told to find an exit and leave the building; it chose roughly the right direction, then walked the robot into a clear glass wall.
6. Fleet data is the compounding moat
Figure has organized the stack around acquiring high-quality, diverse pre-training and post-training data, training a common model and deploying new weights to the fleet. Adcock’s current view is that extending known capabilities to tasks such as towel folding and dishes is primarily a data problem—difficult data, but not necessarily another robot redesign.
There is no separate dishes network, logistics network or downloadable library of motions. Figure reports positive transfer as tasks are combined: added experience improves generalization even when it comes from another segment, echoing Blundin’s analogy that learning piano might make someone “a slightly better soccer player.”
The economic asymmetry is fleet-wide learning. Human skills disappear or must be taught person by person; at Figure, “once one robot learns how to do a task, every robot in the fleet knows it.” That cumulative dataset is both the principal asset and Adcock’s reason for expecting only a few scaled providers.
7. Most humanoid spectacle fails Adcock’s autonomy test
Adcock’s sharpest industry criticism targeted teleoperation presented as autonomy. He compared it with selling a supposed self-driving car while “there was a guy in Tennessee driving it,” arguing that weak hardware can still produce compelling video when a hidden human supplies every decision.
Open-loop martial arts and backflips rank little higher in his hierarchy. A person can perform the motion in a capture suit, after which a small reinforcement-learned policy replays it blind; Adcock said such models may have only about a million parameters and can be trained with open-source code on one desktop GPU.
Diamandis pushed back that kung-fu footage remains fascinating and frightening regardless of its control method. Adcock’s answer was technical rather than aesthetic: replay does not perceive disturbances or reason about the scene, whereas useful robots need closed-loop responses around 200 hertz—“100,000 times” harder, in his rough comparison.
His minimum evidence standard is uncut neural control without a remote operator; he claimed almost no other humanoid footage exceeds one continuous minute. The ultimate benchmark is harder still: drop a robot into an unseen Airbnb and have it complete days of useful work. “We’re, like, so far from that.”
8. Package work supplies the episode’s hardest reliability evidence
The logistics policy runs entirely through neural networks and handles packages at what Adcock called human speed. It separates parcels, finds each barcode, rotates and places the package, and sometimes pats it flat so a scanner under the conveyor can read it.
Figure’s latest reported run produced “one error over 67 hours,” across multiple robots. Blundin emphasized that an operation occurs every second or two, making the run more informative than a short polished demonstration, although the episode did not define the error or total action count.
A six-month Figure 02 deployment at BMW provided a different test: robots operated every workday and proved the system could function commercially. Adcock’s candid retrospective was “80% of the things we got right and 20%…wrong”; the working architecture was too brute-force to replicate across 100,000 units, prompting the Helix 2 redesign.
9. Generalization must precede mass deployment
Adcock rejected manufacturing output as today’s primary scorecard: “What’s impressive today is not manufacturing.” In his estimate, general robotics might be solved with only 100 machines; producing 100,000 units that require operators or replay fixed trajectories merely scales an incomplete product.
The differentiating demonstration would be ten robots entering unfamiliar places and doing useful work, not a giant fleet inside one controlled installation. Long-horizon generalization must therefore be solved before volume has full economic value—“If you don’t solve that, none of this matters.”
Manufacturing still has to advance in parallel because high-rate assembly and robot design require repeated learning cycles. Figure is therefore building production capability while progressing through Adcock’s “level bosses”: short neural episodes, room-scale work, unseen environments, multi-day reliability and eventually broad general purpose.
10. Consolidation leaves only a few global winners
Diamandis compared today’s field—more than 150 Chinese robotics companies by one cited report and perhaps ten serious US players—to the hundreds of early car and tire makers. Asked for an endpoint, Adcock answered “far, far less than 10” humanoid groups globally.
Adcock praised China’s talent, entrepreneurial intensity and hardware output, yet said Figure has seen little closed-loop AI control from Chinese systems. Asked who genuinely threatens Figure, he answered “certainly China” collectively and said he does not currently see another comparable competitive threat.
That competitive assessment did not become a nationalist thesis. Adcock rejected the media framing of an inevitable US-China battle, saying visits to China feel collaborative, like “everybody’s team human” and “team humanity,” even as Figure moves most of its direct supply chain elsewhere.
Every major technology company will nevertheless enter, in his view, because human labor represents a little under half of GDP and the hosts framed the opportunity at $50 trillion. The catch is execution: Adcock put humanoids near rocket-level difficulty and harder than the electric aircraft he built at Archer.
11. Vertical integration is capability architecture, not just cost control
Figure tried to source motors, hands and other systems in its early days and concluded that the technology readiness was too low. A vendor failure in communications, power, sensing, thermals, firmware or reliability leaves the robot maker waiting “or you die,” so Figure now designs core components and performs final assembly itself.
The resulting iteration compressed cost as well as capability. Blundin described roughly a 90% manufacturing-cost reduction from Figure 02 to Figure 03; in a later exchange, Adcock said the Figure 01-to-03 reduction was “about the same” as that comparison and cited machine parts and tooling, without giving a separate exact percentage for Figure 01 to Figure 02.
Off-the-shelf robots lack the necessary sensors, thermal envelope, power and onboard compute. Adcock described teams compensating with giant backpacks, separate batteries, wires and overclocked processors—useful for hobby demonstrations but akin to buying a rocket and fastening a second stage onto its side.
Despite that integration, Adcock forecast that by summer Figure would have “almost none” of its supply chain remaining in China. He framed this as an operational transition, not endorsement of geopolitical hostility.
12. Figure is industrializing production before autonomy is finished
At the production operation called “Baku” and “Bacu” in different transcript passages, and identified during the tour as BotQ, Figure is trying to reach a near-term pace of one robot every 30 minutes. Adcock committed to placing Figure robots on those assembly lines during 2026, then gradually displacing human work through more humanoids and conventional high-volume automation.
The present facility can accommodate four lines, each rated around 12,000 units annually, for slightly under 50,000 at full utilization. Figure is currently building thousands, then intends to move through tens of thousands and hundreds of thousands before attempting millions.
Adcock expects the current site to look low-volume within five or ten years. Figure is already “spinning up resources” for future facilities capable of millions, but his sequencing remains explicit: learn each manufacturing scale rather than pretending today’s prototype line can jump directly to planetary demand.
13. Early customers are laboratories for a leased workforce
Figure retired Figure 02 at the end of the prior year and intends to deploy Figure 03 across multiple signed industrial and commercial customers during 2026. Adcock said the company knows the targeted geographies, tasks and schedules, while 50-to-100 customer discussions have generated enough demand for the next two or three years.
The preferred commercial structure is leasing—Adcock’s deliberately provocative line was, “Humans are leased, so we lease humanoids.” He remained open to sales because the larger objective is distribution: getting enough robots into daily work to improve products and operational competence.
The Grid, a newly opened test facility, will eventually contain roughly 250-to-300 robots operating around the clock across model homes and commercial settings. A second-story mission-control room receives each robot’s video and telemetry, allowing the team to observe both the fleet and the environment through the machines’ own sensors.
Real deployments expose work that demos omit: fleet operations, safety, maintenance, repair and facility integration. BMW taught Figure these disciplines; the Grid is intended to compress that learning before thousands of customers depend on 24/7 service.
14. Onboard inference makes connectivity and battery less constraining
Blundin contrasted ordinary training GPUs with dedicated inference hardware that is neither an H100 nor a GB300, inferring it might be 10-to-100 times cheaper and faster. Adcock confirmed the important point—not the numerical estimate—that fast policy inference runs onboard without consuming the robot’s entire power budget.
Figure 03 includes Wi-Fi, 5G through an eSIM and Bluetooth; users can even text the robot. Persistent connectivity is desirable for fleet functions, but autonomy cannot depend on it: latency or a dropped network must not brick the machine in the middle of physical work.
A full charge supports roughly four-to-five hours of operation from an approximately two-kilowatt-hour pack. Inductive charging through the feet runs near two kilowatts, yielding about an hour to recharge; thin charging mats can sit by a conveyor or kitchen, enabling opportunistic charging rather than an all-day battery.
15. Hark extends Adcock’s “synthetic human” thesis into digital work
Asked whether AGI requires embodiment, Adcock did not give a clean yes-or-no answer. His definition spans both domains: an intelligent system should reason, remember, communicate and “touch the world both digitally and physically,” rather than begin each chat as an “advanced Google search engine.”
Hark, his recently founded AI lab, is pursuing that digital half. In one example, a single prompt requested a CAD monster truck for his son; the model found and installed a CAD package, learned relevant parameters and produced a clean-sheet design in under an hour while operating tools like a human.
Adcock thinks frontier labs are chasing an overly abstract form of reasoning and copying one another instead of building persistent multimodal agents. The system the hosts called Claude Bot, later Malt Bot, illustrated the product overhang: simple Markdown instructions, tools, MCP and APIs around a model such as Opus could already produce “magical things.”
Figure’s physical training investment is similarly rising: 3,000 B200s were going live for pre-training, with a larger future allocation planned. Across Hark and Figure, Adcock predicted the next 12-to-18 months could deliver “the largest AI transformation we’ve ever seen.”
16. The home robot is being framed as a persistent social agent
Figure’s destination is “a human in a bodysuit” that accepts language, applies common-sense reasoning, remembers context and performs ordinary work. Although speech, memory, perception and physics can be described as components, Adcock believes they ultimately converge into one pre-trained omni model.
Personality and emotional awareness are becoming product requirements rather than cosmetic extras. Adcock wants a robot that notices when his children return from school sad, remembers their history and responds with enough EQ to talk with them—not merely an appliance waiting for the next instruction.
Elder care is personal for him: his parents have operated senior housing in the Midwest for about 15 years. He wants humanoids to help people “age in place” at home, combining household work, observation and companionship rather than forcing every care need into an assisted-living facility.
That breadth supports the general-purpose form factor. Humans perform billions or trillions of distinct activities, and one shared model can learn across them; bespoke pipe-cleaning, mining or surgical machines may remain, but Adcock expects humanoids to dominate the plurality while specialized robots stay niche and expensive.
17. Surgical dexterity may arrive before surgical intelligence
Adcock said he felt “pretty confident” that by the end of 2026 Figure’s hardware could perform most physical motions a surgeon performs. His claim was conditional and narrower than autonomous surgery: depending on the procedure, a teleoperator might use the system for real surgery, while the medical “brain” would still require far higher performance.
This is where he defended teleoperation after criticizing it as a marketed product. Teleop is an excellent hardware test and data source: if payload, range of motion or dexterity prevents a human operator from completing a motion, a learned policy cannot rescue it; “if you can teleop, you can learn it” once sufficient data exists.
Figure is already augmenting human-like sensing. Palm cameras improve blind reaches, tactile sensors measure fingertip contact, torso cameras watch occluded feet and rear cameras expand awareness; Adcock was also open to infrared, ultraviolet and other modalities that could eventually exceed human perception.
18. Home rollout is gated by interventions and a newborn-level safety bar
Figure can already perform pockets of dishes, laundry and kitchen work, but Adcock wants days or weeks of connected behavior in a home the robot has never seen. “I don’t want to ship slop” was his answer to demands for a precise consumer launch date.
His best estimate was that by the end of 2026 Figure could place a robot in an unseen home for fairly long-horizon work. The key metric then becomes human interventions: once an hour, once a day, once a week or once a month. Initial user-home shipments could follow the next year if that curve improves.
Scale would proceed iteratively: one successful home, then ten, 100, 1,000, 10,000, 100,000 and eventually ten million. Adcock acknowledged that competitors might advance faster and even said generality “could happen in a couple months,” but treated a staged deployment and feedback cycle as unavoidable.
Asked when he would trust Figure to hold his newborn, Adcock said, “We’re not there now.” The release bar is free, fully autonomous operation around his own children, backed by redundant real-time safety architecture and an accumulated safety record—not supervised visits where engineers “babysit it.”
19. Safety keeps Figure’s model and hardware inseparable
Adcock separated semantic safety from intrinsic safety. The model must understand why knocking over a candle or boiling pot is dangerous, while the machine itself must remain safe around people, pets and animals even when software or components fail.
Privacy and cybersecurity form another layer because household robots will continuously perceive intimate spaces. Figure has an in-house cybersecurity team spanning product, commercial and corporate systems; Adcock emphasized disclosure of what is collected, where it goes, encryption and keeping the information private.
Figure is also developing foundational behavioral rules for non-volatile memory “at the chip level.” Adcock said the company has its own variation on Asimov’s laws but declined to disclose it; safety, privacy, reliability, maintenance, fleet operations and financing all need to be packaged before mass deployment.
Blundin asked whether Figure might franchise Helix to other robot makers. Adcock’s answer was “No”: without ownership of the sensors, actuators and failure modes, Figure could not guarantee safety. “We have a fiduciary duty to our civilization to build really safe humanoid robots at scale.”
20. Tens of billions of robots imply a financing market as large as the product
Adcock expects manufacturing volume eventually to push humanoids toward $10,000-$20,000. Diamandis—not Figure—translated a $20,000 machine into a hypothetical lease near $300 monthly, $10 daily or roughly 40 cents hourly, arguing that this price would expand household demand far beyond one robot.
If all goes well, Adcock sees one humanoid per person plus roughly five-to-seven billion, perhaps ten billion, in the commercial workforce—“tens of billions” overall. At $20,000 each, however, one billion robots alone represent $20 trillion of working capital; he pointed to trillion-dollar vehicle-leasing and credit-receivables markets as financing precedents.
Two prerequisites dominate that scale equation: a neural model that generalizes and robots building robots. Adcock hopes that within 24 months “all the robots will build all the robots,” while manufacturing software and lines are being designed for humanoids to assemble successors and remove them from the line.
Diamandis asked whether individuals will own robots that earn income or whether hyperscalers capture the surplus. Adcock answered that Figure will sell robots at scale and users will be able to deploy them for whatever work they choose; his larger claim was abundance through ubiquitous goods and services. “It’s gonna feel like 2080 up in here.”
Verification Notes
- The transcript alternates between “Baku” and “Bacu” for the production lines and separately identifies the toured operation as “BotQ”; the digest does not resolve those names.
Full transcript
Peter H. Diamandis
I am blown away by how far you've come.
The things that you can do with neural nets now completely blow my mind. Year to year, the whole business looks completely different.
It's amazing to me how you accumulate data, and the data becomes this incredible barrier to entry—
Yeah.
This incredible asset.
The one thing that's important here is that once one robot learns how to do a task—
Peter H. Diamandis
Yes. Everybody learns.
Every robot in the fleet knows it. Humans don't operate like this.
Peter H. Diamandis
Yeah. When do we start seeing robots building robots?
We will put robots on our Bakku lines this year. Listen, this is going to be the largest economy in the world. It's going to be a super impactful business. It'll lead to ubiquitous goods and services for anybody, an age of abundance, and it's going to be a super fun business, too. It's going to build a sci-fi future we all want.
What you're seeing is that every major group in the world will get into this space. You have to. You have no choice.
Peter H. Diamandis
When are we going to see the first Figure in a customer's home?
My best guess is—
Now, that's a moonshot, ladies and gentlemen.
Peter H. Diamandis
So Dave and I are in San Jose at Figure headquarters.
Mm-hmm.
Peter H. Diamandis
We just did a podcast with our extraordinary friend Brett Adcock.
Over here.
Peter H. Diamandis
Yeah, Figure 01.
This is the original.
Peter H. Diamandis
Yeah.
Still somewhat functional.
Peter H. Diamandis
Yeah. It ran the first large language model, the first neural net.
They built it in under a year. Brett actually was screwing these things together himself, and it was all about gathering telemetric data so they could build this.
Peter H. Diamandis
Here's Figure 02. Much more beautiful.
Mm-hmm.
Peter H. Diamandis
Much more functional, running neural nets across the board, dumping all the C++. Can you live long and prosper?
Yeah.
Peter H. Diamandis
Hi.
Hi.
Peter H. Diamandis
And here we go with Figure 03, the workhorse right now. We just did a tour. I mean, we probably saw 100 of these walking through the hallways, on test stands, cleaning dishes.
Brett, what's been fun is they added a flexible toe, too, so it can go down like this.
Peter H. Diamandis
Yeah.
Before, it had this clunky foot here.
Peter H. Diamandis
And Figure 03 has the palm camera.
Palm cam?
Peter H. Diamandis
Yeah.
They cut about 30 pounds off the weight and 90% of—
Peter H. Diamandis
Of the cost.
Manufacturing cost.
Peter H. Diamandis
Wow, crazy.
Yeah.
Peter H. Diamandis
Amazing. Yeah.
It's all front-loaded.
Peter H. Diamandis
It's the perfect height between the two of us.
Yeah.
1. Figure Rebuilds Its Robot Stack
Peter H. Diamandis
I'm here at Figure headquarters with Brett Adcock and DB2. Brett, it's been about 18 months since we did a podcast on Moonshots together. I am blown away by how far you've come.
18 months in AI time, that's like a decade.
Dude, welcome to Figure headquarters. What do you think?
Peter H. Diamandis
Yeah, it's extraordinary.
Holy crap.
Peter H. Diamandis
I mean, just—
Wow.
Peter H. Diamandis
Describe it. We just went on a tour. You've got about 300,000 square feet, with 400,000 square feet under development here. I mean, there are Figure 03 robots walking down the halls. There are fully autonomous robots, I guess, running Helix 2. You just released Helix 2 today.
Today.
Peter H. Diamandis
I got it while I was flying up here. We have these robots doing everything from kitchen tasks to packages to different types of manufacturing. I mean, how many robots do you think we saw? Seriously.
I wasn't counting. Hundreds, maybe—not a thousand.
Peter H. Diamandis
Yeah.
Hundreds.
Peter H. Diamandis
At least a hundred or so.
Yeah. Well, there are a lot of partial robots out there, too, so it's hard to—
Peter H. Diamandis
Picking up Figure heads. That was fun.
How many hands do you think we saw? There were many more hands than there were robots—
Peter H. Diamandis
The hand line, the head line, the torso line.
Actually, picking up the head was the most surreal—
Peter H. Diamandis
This is where the pelvis is made.
Part.
Peter H. Diamandis
Yeah, for sure.
Yes.
Peter H. Diamandis
I still remember, during my first visit with you—full disclosure, my venture fund is invested in 2 of your earlier rounds. I'm super proud of the progress that you've made. I still remember your Figure 01 putting a Keurig cup in a coffee maker, and that was a big deal because it was done with neural nets and not C++.
That honestly was a big inflection point for us. I think a few things we needed to really run down were: Can you build an electric humanoid that's low-cost and capable like a human, just on the hardware side of things? The second thing is, can you figure out a way to not code your way out of this problem? How do we use a neural net to learn those human-type representations—
Peter H. Diamandis
Yep.
—and then do tasks? When we were doing the Keurig task, it was basically a bimanual neural net running on the robot, which has now evolved into Helix. It was able to do the whole thing. It was a smaller task, but it was a few minutes long: picking up the Keurig cup, opening the coffee maker, putting it in, and running it.
It was the first time we saw a true instance of neural nets really working on a bimanual humanoid robot.
Peter H. Diamandis
Yeah.
That was when we were like, “Okay, we have to just go all in on neural nets. The whole stack needs to be neural nets to make this work.” That started basically 2 years ago now.
Peter H. Diamandis
That's crazy.
And then you guys saw Helix 2 today—
Peter H. Diamandis
Yes.
—which is basically the best release we've ever had.
Peter H. Diamandis
So we'll run a clip of Helix 2 while we're describing it, because what we saw was Figure 03 running Helix 2—
Mm-hmm.
Peter H. Diamandis
—in full autonomy—
Mm-hmm.
Peter H. Diamandis
—going into the dishwasher, picking stuff up, and putting it away, not preprogrammed.
Yeah.
Peter H. Diamandis
And I loved the human elements of it, like using its hip to close something and its foot to raise the dishwasher.
That's the neural net difference, though. You get unexpected behavior, both good and bad—
Yeah.
—but things you could never code up.
You could never code this.
Your career went from software company to VTOL company. Now this has got to be the first neural net platform.
Yeah. The things that you can do with neural nets now completely blow my mind versus code. We could never have done a quarter of the stuff that you saw today with the whole body, with manipulation, with things that—you know, there's only so far you can really push code heuristics. A humanoid robot using only code is just a dead end.
Yeah.
It's just not going to work.
Yeah.
Yeah.
Yeah, it's amazing to me how you accumulate data, and the data becomes this incredible barrier to entry—
Yeah.
—this incredible asset. If you were writing all this in C code, that C code would have millions, hundreds of millions of dollars invested in it—
Yeah.
—and you would not want to mess it up.
Yeah.
With the neural net, you can say, “Look. Hey, guys. Retrain it from scratch.”
Yeah.
Right off the bat. It's just a completely different approach. So that's why people are way underpredicting how important—
Yeah.
—or how quickly this is going to evolve, because it's a completely different paradigm.
Well, we've lived through it. Maybe a year or 2 ago, we had several hundred thousand lines of C++ code.
Several hundred thousand.
Handwritten code.
Probably $100 a line to write it.
Yeah. Very expensive, very hard to test, and very hard to get out reliably.
Yeah.
Also hard to model all the different behaviors that we would need to test—
Yeah.
—and then we removed a majority of all that in Helix 1, where we still had a lot of lower-body control being run in basically the control stack in C++.
Yeah.
And then today, we removed the remaining 109,000 lines of C++.
All neural nets.
All neural nets today. That's the full body. That took it from being able to do really good tabletop manipulation, like you saw with the Keurig coffee, to the work we do with logistics. All of that's being done in neural nets.
We've been showing amazing progress there. But getting the whole body out of there and being able to move dynamically through a scene while manipulating and planning is just a whole other challenge. We basically spent the greater part of a year refactoring the Helix architecture to enable this to work. You're talking about now moving through space like a human.
Yeah.
Having control of the full body, with all joints.
Eye, hand, foot, and leg coordination. Everything is—
Sensor data in—
Yeah.
Cameras, tactile. We have palm cameras.
Yeah.
Basically doing inference on board the robot, fully embedded, and then being able to output torques into the motors and do that at a few hundred hertz—
Yeah.
—in terms of planning and control, and do that reliably on very difficult tasks. These are bimanual tasks where it's grabbing and holding things, planning, moving the body, getting things out of the way, making errors and replanning—
Mm-hmm.
—and fixing this, all done with the neural net now, end to end, over a pretty long horizon. For us, it's kind of room-scale autonomy, so we can now finish the whole room—
Yeah.
—which is important. Next, we're going to graduate to basically the full house.
See, that's one of the things that's really obvious when you're walking around looking at what everybody's doing and working on. You visualize a robot company having lots of people working on microcode, actuators, batteries, or whatever, but there's just a huge number of people out there at workstations, and they must be working on the neural nets. It's just got to be such a dominant part of what makes the thing actually look and feel human. The motions are so smooth.
And everybody, when they think about the history of robotics, they kind of chart these line charts, but it's not like that. It's a disruptive change from dropping that last 100,000, 105,000 lines of C code to moving to an all self-organizing neural approach. Completely different future.
It does. We make these technology progress steps, and I think it's been very apparent here. Every year to year, the whole business looks completely different.
Yeah.
A large part of that is trying to get the hardware, hands, and all this stuff in a good spot.
Yeah.
And then be able to basically have more range of motion, speed, and torque like a human. We're all in on neural nets, so it's been, what is the right data set for that for pre-training and post-training? Do we have the right training cluster? Do we have the right models?
Yeah.
And then deploying those really well on the same humanoid hardware. That's a full loop.
Yeah.
We've actually designed Figure 03 to—if you say what the guiding principle of Figure 03 is more than anything else, it was just designing for Helix. It's like, how do we give Helix a body?
So counterintuitive.
Everything: the feet, hands, head.
So build around the neural net.
We really looked at the neural net and said, "How do we fit this into a humanoid robot, and what are the best sensors? How should it run? What does the operating system look like? Middleware, firmware, embedded software?" All of it is encapsulated in this view that we need to go all in on neural nets and do human-like work.
I'm about to release the 2026 version of my Humanoids Metatrend report. It's a deep dive looking at 100 different robots in development right now, a deep dive into 10 of them, including Figure. 150 pages. You can check it out at Substack for my paid subscribers. Anyway, super pumped. This is a field that's moving at exponential, hyper-exponential speeds.
2. Figure Builds Beyond OpenAI
Peter H. Diamandis
So in the beginning, you had partnered with OpenAI on software, and you made a departure from OpenAI. I mean, I guess, are any of—
It wasn't quite accurate, but—
Peter H. Diamandis
Okay. Well, you can—
It's a story.
Peter H. Diamandis
You can correct it.
I think I met Sam and the OpenAI team, and they were just really interested in getting into robotics. In their early master plan, it was to get into basically shipping home robots.
They really wanted to have a very intimate relationship. They ended up leading our Series B along with Microsoft, and we started working on a collaboration agreement to help work on next-generation models for humanoids. We were, and still are, really big on how we language-condition the whole stack.
An LLM, in a lot of ways, is just this world model. It really understands, in the weights, basically what things are and what it should do. It has a lot of good semantic understanding.
We're trying to figure out how to tap that into the humanoid. How do we learn from this at scale and use some of those representations? The partnership just didn't work. Our team just ran circles around them for basically the better part of a year, and it came to a point where we were doing all the work ourselves internally.
We had a whole team here, a lot from some of the best labs in the world, and we were putting out work after work. The Keurig coffee stuff was done by us.
Peter H. Diamandis
Yeah.
All this stuff was done internally.
Peter H. Diamandis
Yeah.
At some point, it just didn't make sense to train other folks on how we built AI models internally for embedded systems like a humanoid.
Peter H. Diamandis
Did it turn out that the LLM matters at all in the physical world? You could start with an open-source LLM and tune it?
What is it, like a VLA—a vision-language-action model you're building?
Yeah. We basically want to take the semantic grounding, then a VLM—
Peter H. Diamandis
Yeah, like the common sense, you know.
—the vision that it has, yeah.
Peter H. Diamandis
Yeah.
What do we understand from this? Which we have in Helix today is super critical. But getting to a point where we can understand physics in the robot and have it really be able to plan and reason at fast dynamic speeds was something that nobody in the world had ever really done before.
Peter H. Diamandis
Mm-hmm. Right.
I think that's the work that we've been excelling at and that we love: how do we get it to understand physics?
Peter H. Diamandis
I think most of our audience probably knows this, but just to rewind the tape, the LLMs—GPT-2 and GPT-3—were built entirely on text data scraped right off the internet. And then they supplemented that with a ton of other data, also in text form.
That creates this machine that has tremendous amounts of common sense. If you ask it, "Hey, do you know how to play soccer?" it says, "Yeah, of course I do." But then you try to install it in an actual physical moving machine, and it has no idea what it's actually doing.
Well, yeah. We've designed it to touch everything in the world. We have this really high-dimensional robot that has 40-plus degrees of freedom. On the surface, the math around this is—just the dimensionality is really high. You have 40 motors, and they can all spin 360 degrees.
Peter H. Diamandis
Yeah.
So the amount of states the robot can be in, like positions, is 360 to the power of 40. There are more states of the humanoid than atoms in the universe.
Peter H. Diamandis
That's a lot. Yeah.
Yeah, exactly.
Peter H. Diamandis
You're not going to simulate those one by one.
Yeah, exactly. So the question is, then, I need to understand these fine contact dynamics. I need to grab this water bottle. Where do I position my elbow, pelvis, torso, head...
Fingertips?
Yeah.
How do I plan to grab this? How do I put pressure on it?
Yeah.
How do I understand those representations really well, from observations into the actions I’m doing at test time? This is not an LLM.
Yeah.
The LLM knows none of this.
Yeah.
Peter H. Diamandis
No.
The LLM knows this is a water bottle. It probably knows that I need to grab it from the side. But all this implied physics that we need to deal with here—we just have to train models to do that.
It’s actually kind of weird because it thinks it knows how to do it, too. The LLMs feel like they can do things intuitively, and then they completely fail.
Hmm.
But—
We’ve done this. You can zero-shot the LLMs inside a robot. We do it; we still do it even actively.
Oh, really?
They just can’t do anything.
Just for fun, just to watch them fall over.
Yeah. I’m interested in this. A project I’ve been doing is: can we just zero-shot it? I’m working on a new AI lab that I founded recently called Hark, and we have a new AI model that is completely incredible.
Wait, you founded a new AI lab? Rewind the tape here. What?
Yeah.
It’s called Hark?
I founded a new AI lab.
I sent you this.
Did you?
I did. I sent—
Hark?
Yeah.
Yeah. We have a new AI lab we’re working on.
Well, I’ll send it to you again.
We have some new AI models, and we actually put one of them into the Figure robot this month. I said, “Okay, let’s just zero-shot this. Let’s give the LLM—or rather, the model—access to basic commands.” This is a multimodal model. Can we give it acceleration and X, Y coordinates for navigation, basically like a joystick? Can we give it a digital joystick?
Yeah.
I asked it to find the exit sign and get out of the building.
Mm-hmm.
Unfortunately, it was going in the right direction and ran into a clear glass wall.
Peter H. Diamandis
Well, kids do that, too.
Yeah, exactly.
Peter H. Diamandis
I did that when I was a kid.
We’ve really stress-tested this. It just doesn’t work. You’re missing so much world understanding—what’s really happening and how to move my body. We think it’s pretty simple to grab an object with a stationary robot, but the robots we have for humanoids are moving.
Yeah.
The pelvis, head, torso, hands, and arms are all moving. When you’re reaching out to grab something off a table, your pelvis is moving backward. It’s very difficult to command a very high-dimensional robot.
Peter H. Diamandis
Robotic physiology.
Yeah.
3. The Humanoid Market Consolidates
Peter H. Diamandis
Over in China this year, some government employees said, “We’ve got a robot bubble.” I don’t know if you saw that article that came out. We have 150-plus robot companies in China, and there’s a lot going on there. In the US, I would say maybe there are 10 serious players.
Mm-hmm.
Peter H. Diamandis
I mean, 2 or 3 who are extremely serious, including Figure. But there are a lot of potential humanoid robot companies. I was just at CES and saw a humanoid robot explosion.
Yeah.
Peter H. Diamandis
And then as many or more hand companies, which is interesting.
Yeah.
Peter H. Diamandis
I go back to the early 1900s, when there were about 250 car companies and 200 or 300 tire companies. Then this massive consolidation occurs, and GM, Chrysler, and Ford buy and consolidate. What do you think is going to happen with all the robot companies today?
I think it happens in every industry like this, especially in deep tech. These will all consolidate down to a few groups globally.
Peter H. Diamandis
Do you have a guess? Is it a triopoly? That’s the right description. Is it more than 10, less than 10?
It always seems—
Far, far less than 10.
Peter H. Diamandis
Far less than 10.
Yeah.
Peter H. Diamandis
Globally.
Globally.
Peter H. Diamandis
Yeah.
It always seems, in the US anyway, to settle down to 2, 3, or 4, but the borders are not obvious. With cars, a car’s a car, right? Then you had cars and trucks, and those were kind of separate for a while. But—
Peter H. Diamandis
You also have different designs. “I want the plush interior—”
Right.
Peter H. Diamandis
“I want the sportster.” I mean, and that, I think—
Mm.
Peter H. Diamandis
I wonder, are robots going to be differentiated by their vertical application or their personalities?
Exactly. There’s so much more variety possible in robotics.
Yeah. I think everybody is taking for granted how difficult this is. You have to go out and build pretty novel, very difficult hardware.
Yeah.
It needs to be relatively cheap. Then you have to figure out how to make neural networks work on it, make neural networks work on it at scale, and manufacture at scale. Then you have to get these products out reliably, with all of them working every day without any human intervention.
Yeah.
We talked a lot about how we’re doing K-Cup coffee work. I haven’t seen a single humanoid in the world do that, or be able to do that today, globally, and it’s been 2 years.
Yeah.
4. Teleoperation Is Not Autonomy
Peter H. Diamandis
By the way, a lot of the video we see is actually teleoperation. I wonder if people realize that. A lot of the robot companies are teleoperated rather than fully autonomous. What we saw just walking around here was a 4-minute-long, fully autonomous operation on Helix 2, right?
I’ve built a lot of businesses in my day, and I’ve never seen so many companies with a human in the background commanding the robot and putting out updates. I’ve just never seen it. When I first started Figure, stuff was coming out, but now it’s every week: somebody is teleoperating a robot and putting out a video.
It would be the equivalent of having a self-driving car company with a guy in Tennessee driving it, while marketing it as if there were no humans in it.
Yeah, yeah.
It’s self-driving. We’re putting out teasers. In a lot of cases now, companies are selling the service. If you want to do this right, you have to believe in neural networks all the way down the stack. You have to build for general purpose.
Peter H. Diamandis
So the parameters that are going to define the successful top 2, 3, or 4 are neural networks and manufacturing—
Okay, I would say what’s—
Peter H. Diamandis
Not surprising.
What’s impressive today is not manufacturing. You could probably solve general robotics with 100 robots. We’re pushing hard on manufacturing, but you can probably solve that. What’s impressive is a full end-to-end robot that is generalizing to an unseen place. You can drop it into an Airbnb—
Yeah.
—and it can do long-horizon work with neural networks. Any long-horizon work in unseen places.
Peter H. Diamandis
What do you define as long horizon? Hours, days?
I would like to see days of work, yeah.
Yeah.
Fully autonomous days of work, at the very least. We’re so far from that. You have robots out there doing karate and jumping, which are preprogrammed, open-loop behaviors.
Yeah.
They’re not impressive. We do that. We’ve done that stuff here—the open-loop behaviors. Any college kid in a dorm room can do this with a robot.
Yeah.
So I think that, plus teleoperation—teleoperation is not impressive. You could build shitty hardware and still teleoperate it and put out videos. That is not hard.
What’s hard is doing full end-to-end neural networks in unseen places, or generalizing to this. If you can solve that, then the next step is figuring out how to get it out at scale. But we are still in the “who can solve general robotics?” phase.
of the humanoid phase, and it's just not impressive if I can build 100,000 robots right now that need teleoperation or can only—
Yeah.
Or can only do open-loop replay. It's just not cool. Right? Your only job is to build 100,000 robots right now.
Yeah.
We have the capital to do it, and we can do it, but what we really want to solve is that I can give you 10 robots and they can go into unseen places and do real, useful work. That's what's going to differentiate everybody.
Peter H. Diamandis
So iterate that until it's right, and then mass-produce.
Yeah. You basically want to be bringing up mass production in parallel, because building high-rate manufacturing for humanoids is going to be super hard, and you're going to have to go through a lot of iterative design processes.
Yeah.
So that's what we're doing now. We're bringing up higher-volume manufacturing as we're learning how to build in true general-purpose-ness. My view is that, if you think about these level bosses that happen—things that will hurt, that you need to graduate to—you need to graduate to doing very short periods of neural networks, which we haven't seen a lot of in the world today. I don't think there's anything over a minute long in the world that's continuously doing neural networks today in humanoid solutions.
Yeah.
Everything's cut. All the films are cut or teleoperated.
Yeah, yeah.
It's pretty crazy. You watch any video, and you want to see it uncut. You want to see it done with neural networks, not teleoperated. You also want to see the stuff that we showed you here in person today, which is running for hours and hours. We run these robots with neural networks.
Peter H. Diamandis
The kung fu videos, whether they're teleoperated or fully autonomous, are actually fascinating and scary when you see them doing that.
But the technology around that is not great. You're basically putting somebody in a mocap suit.
Yeah.
You're having some guy do karate chops or walk around, and then you're running that open loop. You're running it blind. You're just hitting a replay button.
Right.
You can do that with a very simple RL neural net.
Yeah.
You can basically do DeepMimic on this.
Yeah.
It's super simple.
Yeah.
There's open-source code for this that you can run with basically 1 GPU on your desktop, and you can do it with any robot. Every robot has a very tiny amount of computing power. These are single-million-parameter models. They're very small. You don't need a lot of memory, and they're very simple to execute.
Yeah.
What you really want is good closed-loop control, where it's reasoning at around 200 hertz, or 200 times a second.
Sure.
It's dynamically responding to the scene.
Yeah.
That is literally 100,000 times harder than doing open-loop replay.
Peter H. Diamandis
I wonder what the human cycle time is. How many hertz?
Oh, God.
Peter H. Diamandis
Probably less than that.
Much lower than that.
Peter H. Diamandis
I would imagine.
One thing we've seen in our robot is that we can balance on 1 leg better than a human. We just have much better, faster dynamics.
5. Figure Development Accelerates
Peter H. Diamandis
Can we talk about the speed of development here? So, 2025—I'm just trying to imagine. You put out this beautiful post every week on X about the progress in the robotics field and what's going on here at Figure.
And it just constantly—locomotion was a big step forward, excuse the pun, for Figure. Just seeing it walk and then run very naturally, what else was significant in 2025 for you?
I mean, we launched Helix in 2025, about this time last year. We're about a year in now. I think that was highly significant. We basically figured out how to run neural networks on a robot for long periods of time. How do we get the data for it? How do we train models? How do we deploy at test time?
Mm-hmm.
So you watch package logistics. I think I saw it running—
Saw this.
It's been running for days now.
Yep.
It's a neural net all the way down the stack. It's learning how to grab packages, individualize them, find the barcode, and position it down. It'll even pat the package down so the barcode reader below can see it and scan it. It's doing that with high accuracy and at high speed.
High speed, though. That's the part that jumps.
Human speed.
Well, because a lot of what you see in robotics—
Peter H. Diamandis
It's as fast as a human would do it.
Yeah.
Our last run—we had 1 error over 67 hours, over multiple robots.
Over 67 hours.
Crazy. You know what I'm saying?
Yeah. It's doing an operation every second or 2.
Yeah.
So 67 consecutive hours of that is a lot.
It's nuts.
I'd say Helix is a big one, and then Figure 03. Figure 03 is a huge step change for us in hardware.
Peter H. Diamandis
What do you see going into 2026 here? We've got the next 11 and a half months. What are you excited about seeing?
Yeah.
Peter H. Diamandis
What are you excited about?
Our entire roadmap is around Helix 2 now. Helix 2 can go from doing the logistics use case while stationary to walking and moving, and basically do long-horizon, full-body control. We've now integrated all the sensors—the tactile sensors, camera, and palm—into the stack, and we're seeing improvements overall in the policy layer. We're getting better and faster at taking data and running it onboard the robot now.
I wanted to ask you what defines Helix 2, because you're probably incrementally improving the neural net every day.
Yeah.
So what defines it?
We have a couple of big steps. One is that we've integrated a fully learned controller—what we call System Zero—into the robot. The robot has a full-body reinforcement-learned controller in it.
Okay.
Now there's literally no code running on that robot, so it can move its whole body itself using a full learned controller inside Helix. We call it S0.
Has anyone else ever done that before? That's got to be—
There are reinforcement-learned controllers out there. A lot of the karate stuff you see and things like that use them, but nobody has totally integrated one into the whole body for learned manipulation and perception. Nobody's shown that actually working while moving around and doing the things we saw today. I don't even know if anybody has shown it stationary, standing, and doing learned policies. Actually, probably not in the world. So we're getting it integrated into a stack that we can actually use going forward.
I think one of the things we learned at BMW last year—we were there for 6 months, and we deployed our Figure 02 robots every single day—was that, with the stack we had, we got about 80% of the things right and 20% of the things wrong. The things that we got wrong, we didn't want to scale.
Yeah.
It was working. The robot ran every single workday, and it worked. But we learned that we didn't want to ship 100,000 robots with this architecture stack. It was just too hard to scale.
Yep. Mm-hmm.
It was too brute-force.
Yep.
So we've worked for almost a year now on what the ideal architecture is, where we can go out and accumulate large sets of pretraining data—
Put it in the robot, and it can just do this work, and we get generalization from this. That’s what you’re seeing today.
So Helix 1 still had the C++ code in it. What defines Helix 2 is—
Helix 1 had a lower-body controller—
Okay.
—that was still written in C++.
Yeah.
Everything else—the whole upper body—was full neural nets.
Okay.
And so we basically completed the full body now.
Okay.
In doing so, we also did some work at the system level, the System 1 level, where we integrated all the sensor modalities from the hands—
Mm-hmm.
—and the rest of the robot into the stack. For example, we now have tactile sensors in every fingertip that we’re using on Figure 03—
Oh, yeah.
—as well as palm cameras.
Peter H. Diamandis
Mm-hmm.
Yeah.
That’s to understand when we’re occluded and to have a better understanding of how we’re grasping items. We put in a bunch of stuff about picking pills and things out of pill cartridges, where the hand literally occludes them. Your hand is literally in front of the head camera—
Yeah.
—but we still really want to understand where we’re going.
Yeah.
So now, with Helix 2, we basically have a full stack, end to end, with neural nets. We feel confident scaling the pre-training data set into Helix. In fact, I’ll even go as far as to say we designed Helix 2 for the pre-training data set.
Yeah.
And then we designed the robot for Helix 2. We’ve designed everything around data.
Yeah.
How do we get data at scale? If you’re in the neural net game, it’s a data play. It’s about how high-quality and diverse—
Peter H. Diamandis
So it’s experience. It’s just gathered experience—
Yeah, quantified.
Peter H. Diamandis
—in the field, in all kinds of circumstances.
Where can we find—
It’s weird, and I know everybody knows this already, but it’s accumulating, and it never goes away. It’s incredible—
Peter H. Diamandis
Unique data.
—forward progress. You teach somebody how to scuba dive or how to play piano, and they have that knowledge. They live, then they die, and then you have to teach somebody else. This is completely accumulating information.
The reason why I think there will be very few humanoid groups is that the one thing that’s important here is that once one robot learns how to do a task—
Peter H. Diamandis
Yes. They’re ready to learn.
—every robot in the fleet knows it. Humans don’t operate like this.
Yeah. No, it’s—
Peter H. Diamandis
I wish we did.
I watch my kids. I let kids learn how to do stuff—
Peter H. Diamandis
Some people—
—and they just don’t listen, right?
Vulcans do. Vulcans can do it.
I wish we—
Peter H. Diamandis
Melding.
I wish we did.
6. The 2026 Production Push
Peter H. Diamandis
So, 2026 predictions: what’s your boldest prediction for Figure? What are your goals for this year? What do you imagine?
We’re spinning up Baku production in a big way right now for Figure 03.
Peter H. Diamandis
So you said something like a robot every 30 minutes, you expect?
We’re trying to get there in the near term.
Peter H. Diamandis
Amazing.
You guys saw it—we walked through Baku today. What did you guys think of BotQ?
Peter H. Diamandis
Yeah, it’s—
That’s wild.
Peter H. Diamandis
—a lot of humans there.
I wish everyone could see that. I guess it’s all secret. You can’t bring a camera through there.
We haven’t. There’s a lot of IP there, because you see exposed boards—
Yeah. That’s so sad.
—and actuators and stuff.
Peter H. Diamandis
Well, if we can get people with—
It’s cool, right?
Peter H. Diamandis
If we get some footage—
That was so cool.
Peter H. Diamandis
—in video, maybe we can mix it in here. There are a lot of humans—
When do we start seeing robots building robots?
We will put robots on our Bacu lines this year.
Okay.
Phasing humans out of there will be a combination of getting more robots there and doing more high-volume automation over at Bacu.
Okay, so that’s—
Mm.
—the first 2026 objective. Hit that—
We want to scale up robots at Bacu for sure. The second thing is that we want to scale out robots in the industrial commercial workforce.
Yep.
We have multiple clients that we’ve signed. They’re buying or leasing robots from us, and we’re going to get those out at scale in 2026. We know exactly where we’re going geographically, what the use cases are going to be, and the deployment schedules. We want those to be Figure 03s, so we just retired Figure 02s at the end of last year.
Mm-hmm.
Now we’re basically building the arsenal of Figure 03s as we scale up that manufacturing, to get them out into the world and running every day. We like the commercial workforce because it really helps harden our ability to run robots every day.
Yeah.
What we’re here to do is build robots and run them in the world, and they need to run 24/7.
Your ideal customer is who? I know a lot of people would love—
To be frank, we have so much demand—
You do.
—for customers. We’ve talked to 50 to 100 customers or so in the last 6 to 12 months. We really want to be all in with a smaller group of customers and spend time with them, integrate well into their facilities, and do well. We’re still early, right? We don’t have thousands of robots running at these places. We want to get there as fast as we possibly can.
I think we could probably ship an enormous number of robots to the current customers we have now. We see that we’re kind of good for the next 2 or 3 years in terms of demand. We have so much demand. They’re kind of waiting for us to ship at scale.
Leasing versus sale?
We really like the leasing model.
Yeah.
Humans are leased.
Yeah.
We can model it that way. At least these days. They used to be bought way back.
At least these days. You lease humans, so we lease humanoids today. We’re not opposed to selling them. I think what really matters is trying to figure out how to find the right distribution to get robots out at scale.
Yeah.
It’ll really help us get good at what we do. It’s one thing to show a demo or whatever else, but when we had robots at a commercial customer last year, at BMW, it taught us a ton about running them every day, fleet operations, safety, repair, and maintenance. There are a lot of other things that need to come through in the ecosystem that we need to get right.
I’d say the second thing is getting robots out at scale to commercial customers. The last thing—which is arguably the most important for us—is that we want to solve general robotics.
Hmm. Yeah.
We want to solve general robotics. The analogy is that we want to build a human in a bodysuit that you can just talk to, that has common-sense reasoning, that you can communicate with, and that has almost perfect memory of what’s really happening or what’s going on in your life. It can maybe talk to you and almost be your companion.
I mean—
And then go off and do things that an everyday human would want to do. I would expect them to get up to speed on those tasks as fast as or faster than a human can.
Are there 2 different models driving it, then? The VLA model for the body, the physics, and the embodiment, versus an LLM for conversation and memory?
Are you doing both?
We believe this all comes down to one model at the end of the day—one omni model that is trained early in pre-training and helps fuse all this together. You could think of it like we need to have speech, we need to have language-conditioned policies, we need to understand physics really well, and we need to remember things and be able to recall them easily.
We need to have some sort of personality on the robot. I think one thing that you're going to see more and more is we really want to make this robot something you can spend time with.
Yeah.
We've been really focused on getting the core building blocks built, but over the next year or two, I think you'll see us... I think I just want a robot in my home I can talk to—
Sure.
—that can remember things, especially my kids. My kids come home sad from school or something, and I want the robot to understand that.
Yeah.
I want it to have the EQ—the self-awareness to see that and talk to them. I think all this is something we want to spend more time on. We're spending more time on it now internally.
Is there already a big MoE model where, depending on the task you're doing, it'll run different parts of the neural net, or does it need everything?
We have one neural net now that's basically—
One module.
There are no libraries of neural nets that we pull down.
That's interesting.
So there's no dishwashing neural net or logistics neural net, as you saw here.
Yeah, because at scale, if you teach the thing every physical motion, there's a massive number of combinations. The storage is actually dirt cheap.
Yeah.
But the processing is very expensive.
Yeah. Even better—
So—
We've seen positive transfer now with all this data coming in. The robot can generalize better with more information, even if it's the same segmentation.
Yeah.
Yeah.
It's weird. More knowledge is better. It does cross-transfer. Playing piano—
Yeah.
—makes you a slightly better soccer player.
Yeah.
But you don't want to run the whole parameter set for piano playing when you're playing soccer. It's an interesting little hybrid problem there to—
Yeah, you don't want to nuke it. Yeah, for sure.
Yeah.
That's where we try to build the best-in-the-world models here and build a great team that can ultimately deploy robots that are useful. I think showing this type of usefulness—whether it's a lot of the stuff you saw today or a diversity of that—is super important for a humanoid robot. It needs to be able to do everything a human can, which is—
Yeah.
And on the distribution curve, we probably do billions or trillions of unique, very unique things in the world that we talk about.
One of the things you said on our tour that totally tells me we're on the right track is that you're using normal GPUs for training, like everybody, but inference-time compute is on super-fast, dedicated, non-H100, non-GB300 hardware—
Yeah.
—which has got to be at least a factor of 10 or 100 cheaper and faster, which means—
Yeah, it's also running fully onboard.
And it's running fully onboard. Yeah.
So we can basically do very fast inference and policy deployment. Yeah.
And it's also not sucking down the entire power of the robot.
You also have an issue where we've run models offboard the robot, but if we lose communications or have some network latency—
Peter H. Diamandis
So I wanted to go there—
Yeah. You know what I mean? If you lose internet, it's hard to do work, and it's—
Peter H. Diamandis
Yeah.
For most humans, it's—
Peter H. Diamandis
We hit on supply chain, batteries, and comms. On the comms side, do you imagine you're going to be running a 6G network on there besides Wi-Fi? What's going on in batteries these days?
Yeah.
Peter H. Diamandis
Yeah.
From a network perspective, or communications back to the robot, we have Wi-Fi onboard. We have 5G and a SIM card—an eSIM—onboard, so we can... The robot, you can text the robot. You know, we can—
Peter H. Diamandis
You can have James[?].
Yeah.
Yes.
We can have a network outside of a Wi-Fi connection, and then we also have Bluetooth onboard, so it's almost like a walking phone or something like that.
Peter H. Diamandis
Mm-hmm.
Yeah.
We want connection at all times, but ideally you also want the robot to be able to perform work without a connection. So you really want a lot of onboard intelligence, in case you lose internet, so the robot's not bricked. Humans, for the most part, can do work without their cell phone. Not teenagers.
Yeah, that's going to be like that.
Not teenagers. Okay.
Peter H. Diamandis
So, batteries—
Yeah.
Peter H. Diamandis
They've been improving. What's the battery life right now? I love the charging mechanism, by the way. For those who don't know, you're charging basically through your feet.
Through your feet.
Yeah, no connector. You just stand.
Yeah.
Peter H. Diamandis
Yeah, which is kind of like—
You can stand there.
Inductive charging.
It's really cool.
Peter H. Diamandis
What kind of battery life are you getting? What do you expect in 2 or 3 years for battery life? So today, it's what?
We run basically around 4 to 5 hours per full charge in the battery, if we're starting at full battery life.
Peter H. Diamandis
Sure.
And then through full depth of discharge, we can charge wirelessly at about 2 kilowatts through the feet, inductively. We have about a 2-kilowatt-hour battery pack, so it's about an hour or so for a full charge—
Mm-hmm.
Peter H. Diamandis
Hmm.
—on the robot. So we can do 4 or 5 hours on, an hour off.
Peter H. Diamandis
That's great.
Yeah, it's great. I think folks are over-indexing too much on how long the robot can run on a single charge.
Peter H. Diamandis
Mm-hmm. Yeah, I don't expect—
The only thing that's really—
Peter H. Diamandis
I don't expect there are that many tasks you need a robot—
Humans take a few hours in. You go take a little break and do stuff. So I think there's ample time to do opportunistic charging, maybe send another robot in. We can put this little thin mat anywhere in the world. It could be on a conveyor system or wherever else. It could be at home in front of the kitchen, and you can just charge there while doing work.
Peter H. Diamandis
Mm-hmm.
Yeah.
Peter H. Diamandis
By induction.
Which is really cool.
Yeah. You don't have to have any wires or things like that you're pulling from the wall.
Yeah.
Well, I think one of the greatest value-adds you're doing right now is that people are over-indexing on all kinds of weird things because they're physical beings, and they're watching the robot do physical things and saying, "Oh, my God. Can you believe it can sprint now? Oh, my God, it can do a backflip now. Oh, my God, it can do..." And you're like, "Well, it depends on whether you programmed that in C, teleoperated it, or did it actually learn this from all—"
Yeah, I think most of those are open-loop. They're just replay buttons.
Yeah, exactly. It's so hard. So when people say, "How long does it run on one charge of the battery?" you're relating it to your cell phone.
Peter H. Diamandis
Yeah.
But it's not relevant in the—
I think—
—the inflection point we're going to.
Yeah. The summary here is: I need to see real closed-loop control of a robot moving around, touching and moving things like a human would.
Yeah.
That's where the hardest problems all sit, and that's where we've seen this huge wave of humanoid explosion, like you said, out of China and things like this.
Yeah.
But we've seen a very steep drop-off in getting to the next point, which is: show me a minute of the robot doing Keurig or something like that.
Yeah.
Uncut, closed-loop.
Yeah, real-time.
Yeah. You just haven't seen that. And I think you will, and I think there's a lot more levels to go from there. And that took us 2 years to go from a few minutes of tabletop manipulation with neural nets—
To a point where we can do kitchen work—room-scale autonomy. That was 2 years of working 7 days a week. We were here a lot of nights getting there, so it gives you a little sense that you're not going to do that in 6 months from there.
I think there's a lot of hardware, low-level C software, firmware, embedded systems, sensors, neural nets, and data. All of that came together to build this. We couldn't do the same work today on a robot that we could buy off the shelf today.
7. Figure Builds Its Own Supply Chain
Peter H. Diamandis
You vertically integrated and made the choice to vertically integrate, but how much of the supply chain ties back to China?
Hmm.
I think in the next—I think by summer we'll have almost none of our supply chain in China anymore.
Peter H. Diamandis
Amazing.
Anymore.
Peter H. Diamandis
Do you buy into the U.S. versus China AI and robot competition? How do you think about that?
I don't. I spend a decent amount of time in China, and I love it.
Peter H. Diamandis
Yeah.
China is great. I go there and, you know, you're watching TV here in the U.S. and it's just this massive conflict and battle and everything. Then you go to China and everybody's just trying to help and win, trying to work and collaborate, and it feels like a startup incubator.
Peter H. Diamandis
It's just like a one-way—a one-way competition.
It just feels like everybody's Team Human.
Peter H. Diamandis
Yeah.
Team humanity.
Peter H. Diamandis
Yeah.
Go win. It's so great. Then you come back here and you're poisoned with all this stuff online, articles, and television, and it's just not like that when you're boots on the ground and going to do this. It's like, let's go as one and go win.
I love that spirit of trying to progress this technology as a giant lever arm for humanity, to bring abundance for everybody and make it a sci-fi future we all want to live in, which is really exciting.
Peter H. Diamandis
Yeah. Oh, my God. It is. We want to speed-run Star Trek, is what we talk about.
Exactly.
Peter H. Diamandis
I see a figure on the Moon, a figure in orbit—
A figure on the ocean floor.
100%.
The equivalent is that you guys make your own actuators and motors here, and part of that is because you want the exponential-growth effect, but part of that also is that the supply chain just doesn't exist to give you the parts here.
Well, yeah, the improvements—
In China—
I mean, you and I were just talking about the improvements made between Figure 02 and Figure 03—
Yeah.
Because you have all of the ability to iterate in terms of speed and cost. I mean, the numbers that you shared on cost—it was like a 90% reduction?
Yeah.
Is that right?
We reduced costs like crazy on Figure 03.
That's crazy.
It would be great if we could go out and buy motors and plop them in the robot. It doesn't work like that. It would be great if we could go buy hands and just screw them onto the end. It literally doesn't work.
If you go through the engineering work to understand how we do comms, power, sensors, failure cases, thermals, low-level firmware, and embedded software, there's a cost. If something breaks in that equation, or there's a reliability issue, you're left with, hopefully, the vendor fixes it—or you die. It just doesn't work.
Yeah.
None of this stuff—the technology readiness of these things is really low. We would have loved to have gone out and bought all this stuff in the early days. We tried.
Yeah.
We basically failed at all of it. So we thought, “Okay, we need to design it ourselves,” and now we manufacture it. We do all final assembly and everything here.
We do that in some cases because nobody knows how to do it well. We do that a little bit for IP—we really want to control that here and understand what people have access to. We also want to get good at making a lot of robots.
What we need to get good at long term is probably a few things: getting data at scale that can run neural nets, basically doing Helix really well, making a lot of robots, and then getting those things out in the world at scale. A pretty simple equation.
It just feels like—
Yeah.
The journey of getting Figure up and running must have been so much harder than it would have been in China. But then, once you have everything built in-house—all the actuators, training the neural net, and everything in-house—you have a massive advantage versus anything going on in China. If you'd been locked into a supply chain—
It's—
—that only has certain models and makes—
It's not even that. Even if we used an existing supply chain for all this stuff, the robot wouldn't be able to do what you saw today.
Yeah.
It just can't do it. If you go out and buy a humanoid robot off the shelf today, we can't get it to do this.
Yeah.
We've bought robots off the shelf. We've looked at them. You just can't get them to do this work.
Yeah.
They don't have the right sensors. They don't have compute. They don't have thermals. They don't have the right hardware—the hands, the head. All of these are built around our neural net stack. All the bots—
Well, that's a new thing, too. The neural net is incredibly integrated with this specific hardware.
If you watch folks who are trying to buy these robots off the shelf, say, from China, they'll end up retrofitting them with these giant backpacks.
Yeah.
They'll have power, compute, and thermals there. There'll be a wire hanging out. They'll hook that into the back of the robot. It probably has its own local battery. They have to take it and overclock it, and it's just the wrong way of doing this.
Yeah.
It's a hard thing. It's like buying a rocket and saying, “We're going to put stage 2 on the side,” or something like that. It doesn't really work at scale.
Yeah.
It works in the early days for hobby-grade demonstrations and things like that.
Yeah.
But if you really want to do robotics at scale, then you're going to have to go design it yourself.
Looking at the companies coming out of China—Unitary, Engine AI, and so forth—which ones are you most interested in or excited about as friendly competition, if you would?
One thing that's great about China is that there just seems to be an explosion of really great talent and robots coming out the door.
And great entrepreneurial work ethic there, right?
It's awesome. It's great.
Yeah.
Things are just good for humanity, and this needs to happen. I think the thing that we've not seen is any closed-loop AI control from these systems at all.
Yeah.
We've seen a huge lack of that stuff. Usually it's like, “Here are the robots. We'll sell them,” and they're doing a ton of basically open-loop work, looking at—
Yeah, they're hand controllers.
Yeah. Doing that is almost orthogonal work from designing the system the right way for full autonomy.
But if we were to think about who Figure really competes with as our main competition, it's certainly China as a whole.
For manufacturing, I mean—for human, low-cost labor.
I think just for humanoids, we really don't see anybody else besides China as a real competitive threat today.
Fascinating.
Peter H. Diamandis
Rumors about Apple getting into the business? They cut down their car project, and the rumors are that they're heading toward humanoids.
Have you heard that?
We’ve heard this. I’ve been in conversations with every major tech company in the world over the last 12 months.
And then NVIDIA, Google, and even Sam—
Yeah.
Everybody’s making noises about heading into the space.
Meta, Amazon.
Yeah.
Yeah.
Yeah.
Listen, this is going to be the largest economy in the world. Roughly a little under half of GDP is human labor.
Right, $50 trillion.
Yeah. This is the next great place to be, and I think it’s going to be a super impactful business. It’ll lead to ubiquitous goods and services for anybody, an age of abundance, and it’s going to be a super fun business too.
Yeah.
It’s going to build the sci-fi future we all want.
Yeah.
It’s going to feel like—
Yep, it is.
It’s going to feel like 2080 up in here.
Yep.
So what you’re seeing is that every major group in the world will get into this space. You have to. You have no choice.
The major groups being Apple and Microsoft—
Yeah.
Google.
I think every major player that wants to do this.
Yeah.
I think what’s going to be hard is that we’re doing rocket-type difficulty and design here.
Yeah. Mm-hmm.
If Meta were building rockets, you’d be like, “That’d be crazy.”
Yeah.
I would think humanoid robotics is probably up there with rocket design. It’s certainly harder from an engineering perspective than when I built Archer, building electric aircraft, and that was hard.
Yeah.
These were 6,000-pound aircraft with 12 motors and 6 independent battery systems. We built our own control stack and embedded systems. We did all the structural design ourselves, things like this. So I think it’s probably up there with some of the hardest hardware on the planet, and you just have to be all in.
Well, let me ask you about that, because we were talking backstage at Abundance 360 last year, and you had—the basic tech stack had 6 layers of competency. You could probably rattle them off the top of your head, actually.
Is this for Archer or Figure?
This was for robotics prior to neural nets, I guess. So it applied to Archer and Figure.
Sure.
What were they again? It was—
Archer was basically a flying aircraft.
Yeah.
So I basically build electric vertical takeoff and landing aircraft.
Right.
It’s a flying robot, is what I meant. It has battery systems on board.
Yep.
It has electric motors.
Yep.
Electric motors are basically a stator-rotor gearbox. We have a little bit more sensors in our actuators than that, but for the most part, there are encoders and things like that in there.
Okay.
You have basically control software: How do we control this thing and make it move around?
Yep.
In the case of Archer and Figure, it’s a very over-actuated system. Archer has 24 degrees of freedom. We have tilting propellers, and we have pitch on the blades. We have flaps on both the tail and the wing. Figure has over 40 or so degrees of freedom in the system.
Yeah.
You have embedded software on board and sensors.
Okay.
So how do you get the compute, sensors, and embedded software all to talk to each other?
Yeah.
Then you have structures.
Okay.
Those are the core ingredients of a robot or something physically moving through the world.
So then my question is, traditionally, the employee base would be experts in 1, 2, 3, 4, 5, and 6.
Yeah.
They’d be really, really good.
Yeah.
So then you come in and overlay this with Helix, and you’ve got this massive neural network thing. Is that a seventh competency, or is that something that permeates the others? Or did you take all of your microcontroller experts and start training them on neural networks?
The next thing in Archer is: How are you going to plan? You do it through a pilot. My aircraft, Midnight, is a piloted 4-passenger aircraft. So who’s doing the planning?
Yeah.
You have higher-level behaviors in the stack that tell the lower-level control and code what to do.
Yeah.
Here at Figure, it’s been changing over time, but now it’s entirely neural nets—with Helix too.
Hmm.
What is the highest-level behavior telling the rest of the stack what to go do?
Yeah.
Where does that come from? It can come from a human. It can come from a joystick. It can come from an open-loop behavior, which we see, like we talked about before, or it can come from a neural net that’s doing the planning and reasoning.
The kitchen demonstration you guys saw today that we released—what’s telling the robot what to go do next? What’s telling it to pull the rack out of the dishwasher and go grab the cups—not the coffee cups, but the water cups? That’s a neural net making that plan.
Yeah.
In the case of my aircraft at Archer, it’s a pilot determining when to take off, when to hover, and when to transition into full flight.
Yeah.
And then how to descend.
And it’s a different type of neural net.
It’s a human biological neural net.
I guess.
Yeah.
8. Robots Move Into The Home
Peter H. Diamandis
Let’s talk about application layers. We’re seeing your movement into the home, besides the industrial base and such, and healthcare is going to be a big part of this. That includes elder care and helping people stay healthy at home.
By the way, you just came through Fountain? You got through Fountain Life?
Yeah, I did.
Peter H. Diamandis
How was the experience for you?
Thanks for referring me. It was great. I went down to a clinic a couple of weeks ago.
Peter H. Diamandis
Which one? In Orlando?
Orlando.
Peter H. Diamandis
Yeah, headquarters.
I didn’t know what to expect. I’ve done full-body MRIs, CT scans, and blood work before, but when I got there, it was basically a full stack. You know this, but it’s a full stack.
Peter H. Diamandis
Everything measurable about you. 200 gigabytes of data.
Exactly. I spent about 5 hours there, left, and got the download last week. It was unbelievable. What was great about it was that you get a comprehensive understanding of my body and what’s happening, but also somebody there reporting it out and talking me through how we understand it and what to do next.
Peter H. Diamandis
And a plan.
And basically build a plan from there.
Peter H. Diamandis
Yeah. Yeah.
It was great. I actually purchased it as well for my parents. I think it’s just a great gift.
Peter H. Diamandis
Oh, hell yeah.
Yeah.
Peter H. Diamandis
Dave, we need to get you there too.
Why Orlando, and why not somewhere else, like Wyoming or something?
I was on the East Coast, so I popped down to Orlando. It was just easy for me.
Peter H. Diamandis
Yeah, so we’ve got New York, Orlando, Naples, Dallas, Houston opening, Miami, and Los Angeles. Anyway, back to the conversation here.
Yeah.
Peter H. Diamandis
I can imagine this is going to increase the value of health in the home a lot, right? One of my visions of the future is that you’re constantly being monitored for your blood biochemistries—your protein levels, your vitamin levels, and so forth—and that information is being uploaded to Figure in the kitchen, ideally cooking your meals suited to what you need in that moment.
Yeah.
Peter H. Diamandis
And then there’s the whole elder-care side.
Yeah.
Peter H. Diamandis
How do you think about that?
About that.
Growing up, I grew up on a farm in the Midwest, and then my parents got into independent and assisted living about 15 years ago. I kind of grew up around senior care a little bit in my life.
You got into that business?
Yeah. My parents own and operate senior-housing facilities in the Midwest.
Nice.
So wait, they’re still in Illinois?
Yeah, still in the Midwest.
Wikipedia says your hometown has 2,000 people in it. Is that right?
I grew up in Maquoketa, Illinois. I think it was about 1,800 people when I was growing up.
Farm boy.
Peter H. Diamandis
Small.
Yeah, in the middle of nowhere. We had no traffic lights and no fast food. It was a dry town. It was just a whole different world.
Oh, man.
Yeah.
Do they have parades for you when you go back home?
Man, it’s just—
They have robot parades going through the streets.
Yeah.
Peter H. Diamandis
Can you imagine that?
Yeah.
So do you understand the value of a fully autonomous humanoid robot?
We’ve got to figure out how to ship robots into senior care and let people age in place at home. I’m really passionate about that.
Yes. Yes.
It’s hard—
Age in place.
It’s hard to get people to move into assisted-living and care facilities.
Well, how does that work? You’ve sold out 3 years into the future. You can’t make them fast enough to keep up with the demand, and then you’ve got BMW and a bunch of industrial use cases.
Yeah.
But then you’ve got this in-home—
Yeah.
And you’ve got—
Really? Okay. Maybe I’ll level with you on how I think about things. We’ve been spending the last 3½ years—we’re about 3½ years old—trying to figure out what the right recipe is, in the first instance, for what a general-purpose architecture would look like for humanoids. We believe we’ve found it internally, and we understand what that is.
Yeah.
We believe we know how to make robots now and put them out, and we’re going to run them really hard this year.
You showed us—what do you call it? The Grid?
Yeah, the Grid.
Yeah.
The Grid is my favorite place here. We have 4 buildings on campus, and it’s one of them. We have the facility outfitted to expand to hundreds of robots that will run 24/7. We have this little mission command post on the second story, kind of like a 007 situation room, and you can see every robot there.
It’s going to be doing both home and commercial workforce. We’re spinning up right now. The facility just opened this week.
Mm-hmm.
You guys saw it.
Yeah.
It’s squeaky clean, and we’ll start shipping Figure 03s into it this month.
Yeah.
Peter H. Diamandis
So: model homes, model factories, model operations.
Yeah.
Within mission control, you think of watching the robots, but the robots also have their own vision, which transmits back. It’s more like those combat movies where, back at the home base, they’re watching the invasion or whatever. You’re seeing through the eyes of the soldiers.
Yeah.
Peter H. Diamandis
Yeah.
You’ve got all that data coming back into mission control, too. So, if there are 200—how many are in there at any given time? A couple hundred?
250, 300.
250 or 300 robots—
Yeah.
—building a house or doing whatever, and all that video and telemetry comes back into mission control as they do it.
Yeah.
9. Embodiment Defines AGI
Peter H. Diamandis
Do you believe that AGI requires embodiment? There’s a lot of conversation that’s been put forward on that note.
I’m getting the chance right now to spend a lot of time on both physical AI and digital AI at Hark, so kind of both. I think when I talk to AI today or use it, it just feels like it’s so dumb.
Mm-hmm.
It just feels like you’re starting a new chat. You’re basically asking it for knowledge retrieval. It’s an advanced Google search engine.
What I envision is that we want to build the future. We want to build Jarvis—
Yes.
—or we want to build the Jetsons. I want to talk to it.
I want Jarvis so bad.
I want it to talk to me. I want it to reason. I want it to have perfect memory. I want it to be able to touch the world, both digitally and physically. I want to build it to be general-purpose, able to do things for me and reason through things.
Yeah.
We have Hark now designing CAD from scratch. You ask it to build a CAD thing. I asked it to build a monster truck for my son in CAD, and it’s going out, finding a CAD package, installing it, opening it up, and learning how to build CAD and the parameters it needs to look at for building monster trucks. It goes off and does it, and we can do that in under an hour now, fully end to end.
Clean sheet.
Clean sheet from a single prompt, and it’s using tools and computers like a human can. We’re going to give it all the same tools. We’re going to give it all the tools that Figure uses for CAD and FEA, all this different stuff, and it’s going to learn all this.
Was that the inspiration for Hark? The fact that there are a lot of LLMs out there doing a lot of things, but none of them are really connected to CAD, and you have so much experience from your—
My inspiration for Hark is that I feel like all the big frontier labs are chasing this very abstract version of reasoning.
Well, specifically, Anthropic wants to dominate coding and code self-improvement, and then OpenAI wants to dominate—
I want to dominate a sci-fi AI future.
Yeah.
Well, Jarvis. Everyone knows Jarvis.
I want to build Jarvis. I want the smartest person in the world with everybody.
Yeah.
There’s also the von Neumann probe. It’s the idea that these things go out into the solar system and then ultimately out into the galaxy and start making themselves out of raw materials.
Nobody’s doing this.
Yeah.
Everybody’s copying the other frontier lab that’s copying their frontier lab. Nobody’s building true multimodal systems that really can reason and understand, have persistent memory, and go out and touch the world and do things. That’s my version of AGI: I can do what humans can do, and humans are not sitting there giving me Google search answers.
Right. Right.
Which is what we have now.
Right.
It’s terrible. In one aspect, it’s great because this new alien technology dropped on the planet in 2022, and we’re trying to figure out what to do with it. But the other aspect is that there’s so much the models can do now.
There's such an overhang in the product capabilities.
Mm-hmm.
We're understanding that better now at Hark. We're understanding that better now at Figure. I think we're abstractly getting to a place where we're building synthetic humans at scale.
Mm.
These humans can be both digital—they can work on the computer and use tools—and physical. They'll be able to reason with you, talk, have memory, understand you, and go off and do anything a human can.
Yeah.
Peter H. Diamandis
Have you been tracking Claude Bot now, Malt Bot?
Yeah, I've been tracking Claude Bot.
Peter H. Diamandis
Yeah.
It's really cool.
Peter H. Diamandis
Yeah, they renamed it to Malt Bot.
I think it just shows you how complacent a lot of the frontier labs have been.
Peter H. Diamandis
Yeah.
Where you have such incredible—
Peter H. Diamandis
They're—
—capabilities—
Peter H. Diamandis
Capabilities—
—that can be used with a very simple harness and very simple Markdown files, and very simple tools you can give it, on the back of Opus or whatever you're going to use.
Peter H. Diamandis
Yeah.
It can do magical things for the world.
Peter H. Diamandis
Yeah.
And we've had that for a long time now. It wasn't like they went out and built a new AI model for this. They basically just put some harnessing, MCP, and APIs around this, and it basically went out and can basically be your executive assistant.
Peter H. Diamandis
Mm.
It's really awesome.
Peter H. Diamandis
Yeah.
And there's a huge area here to give that to every person in the world and make it easy.
Peter H. Diamandis
Yeah.
We're doing some model development now at Hark that is truly state-of-the-art, I think, and I'm excited about that. We're also doing some of that now in the physical world at Figure, so we have this digital-versus-physical thing that I'm seeing on both. I'm just so excited about this future, even over the next 12 to 18 months. The next 12 to 18 months, I think, will be the largest AI transformation we've ever seen.
Peter H. Diamandis
Yeah.
And getting back to your point about what we do with healthcare and robots, we're going to make a shit ton of robots. We're spinning up resources right now, both at Bachy you're seeing now and future Bachy, to basically be able to make millions of robots.
Peter H. Diamandis
How long before these robots are your physician, your surgeon, able to actually support all the complexity of a medical procedure?
I think from a hardware perspective, in 2026 we'll be able to do—
Peter H. Diamandis
Mm-hmm.
—what surgeons can do.
Peter H. Diamandis
Yeah.
I see no reason we can't do that, given where we're at with our roadmap and things like that with Figure.
Peter H. Diamandis
That's pretty—
Peter H. Diamandis
That's pretty fast.
Yeah, it's pretty fast. I feel pretty confident that by the end of this year, you'll have a hardware system that, if you could teleoperate it or something like that, you could basically be able to do real surgery. It depends what type, but I think most things a doctor could do.
Peter H. Diamandis
And then the AI system is just layering on top of that.
Yeah. Then you have to get the brain to work really well—
Peter H. Diamandis
Yeah.
—at these things, and this has got to work at the highest level of performance.
Peter H. Diamandis
Let me ask you—
But—
Peter H. Diamandis
—federated learning gives you an incredible amount of—
I think we're—
Peter H. Diamandis
—knowledge.
I think we're very close to this working. I think we've already shown that if we can get the right data and the hardware—if the hardware can do it—the simple hack is that if you can teleoperate the robot to do it, we can learn it.
Peter H. Diamandis
Yeah, I mean, that's a really—
And—
Peter H. Diamandis
—important point people need to understand. If you can teleoperate the robot, if the mechanical—
It just—
Peter H. Diamandis
—systems, the motors, and the fidelity can be done.
Yeah. I think we're dumping on teleop, but teleop has a couple of good things. It's a really good testing tool.
Peter H. Diamandis
It proves out the hard—
And it proves out the hardware. If you can't teleoperate it, you're not going to be able to learn it, meaning if there are restrictions in the range of motion or payloads—if you pick up something heavy and the robot can't do it during teleoperation—it's not going to be able to do it in a learned policy.
Peter H. Diamandis
Yeah.
So I think if you can teleop, you can learn it from a hardware perspective. We'll be there in terms of more dexterous-type things we talked about here. What we've already shown is that if we can get the right data for it, we can get the hardware to basically do anything it's capable of.
Peter H. Diamandis
And then you can add infrared, ultraviolet, and all kinds of additional sensors into the system.
Yeah, for sure. We have it now with tactile sensors—the palm camera is a good example. Humans have palm cameras.
Peter H. Diamandis
I would Google that.
We've been seeing a boost in performance.
Peter H. Diamandis
Maybe I do.
A lot of cool things. We're reaching into a cabinet now, and we can use the palm cameras.
Peter H. Diamandis
It totally makes sense. As soon as I heard it on the tour, I was like, "Duh." It totally makes sense. I mean, how many times a day are you reaching? You can't—
We're kind of blind reaching in.
Peter H. Diamandis
—just putting your phone down there to get the camera to look at it.
Yeah.
Peter H. Diamandis
I'm sure we would have evolved an eye right here if it were—
Yeah.
Peter H. Diamandis
—physically possible.
It is an interesting question for you. You've got the cameras in the head, again mirroring a human, and the hands. Why aren't there cameras rear-facing or 360-degree-facing?
Peter H. Diamandis
Maybe there are.
We do.
Peter H. Diamandis
I just bought an amazing drone, the Antigravity drone. Have you seen it? It's the VR headset.
Mm.
Peter H. Diamandis
It's got 360 above, 360 below, backward, and forward.
Yeah.
Peter H. Diamandis
And it's extraordinary. So how do you think about cameras?
We do. We have them on the robot.
Peter H. Diamandis
You do?
Yeah.
They are rear-facing cameras.
Peter H. Diamandis
Okay. I have to ask this question for our Moonshot Mates—
Yeah, if you just—
Peter H. Diamandis
—for some—
—go over and look behind them, they have cameras in the back.
Peter H. Diamandis
Okay.
Yeah.
Do they?
Peter H. Diamandis
I mean, I'm seeing it rotate here—
Yeah.
Peter H. Diamandis
—on here.
Yeah.
Peter H. Diamandis
One of our Moonshot Mates, Salib Alsmal—you might know him. He's one of my co-founders with Ray at Singularity University. He's like—
Nope.
Peter H. Diamandis
"Why in the world are there only 2 hands? Why don't we see robots with 4 hands or 6 hands?"
Yeah.
Peter H. Diamandis
To put that to bed once and for all, personally.
Yeah. We get asked this a lot. It's like, "Why not superhuman?" and all these different things, which is a lot of the questions. My summary to this is that our goal is to be able to do what humans can, and then you want to do it the cheapest and lightest possible way you can. The lighter, the better, for safety. The cheapest is obviously very important. All of those will affect manufacturability and scale.
When you start building things that are better than humans in a lot of ways—if it can run a 3-minute mile, if it can do a backflip, if it's got a bunch of arms—it's going to make the robot really heavy.
Peter H. Diamandis
Yeah.
It's going to make it really costly. It's going to be really hard to manufacture.
Peter H. Diamandis
It's optimization.
And then your question is, okay, when I look at the logistics use case, I don't think you can actually have 4 arms or 6 arms and move any faster. The line is relatively small—it's maybe a meter or so in depth. You have to get a package. The package needs to be roughly in the center of the conveyor system so the scanner below it can scan it and put a label on it.
Peter H. Diamandis
Mm-hmm.
In that case, we basically have another 3 to 5× in terms of speed that we can run the actuators at. The software isn't enabling—
Peter H. Diamandis
Mm-hmm.
—because it doesn't know how to do it yet. So we can run 3 to 5 times faster than what you saw today.
Peter H. Diamandis
Wow.
Because we need the whole body to run.
Peter H. Diamandis
That’s amazing.
Yeah. When we look at it in terms of radians per second, we traditionally look at RPMs.
Peter H. Diamandis
Yeah.
We look at radians per second here. We have another 3 to 5X headroom in the actuators that you’re seeing now.
Peter H. Diamandis
I would love to see a robot do that.
I think the cost of a mistake when you’re unloading the dishwasher—
Yeah.
At the current rate of speed, the cost of a mistake is relatively low. You start running 3 to 5X faster and you—
It’s just like—
If that thing glitches—
I just don’t know if it’s really needed.
Yeah.
You’re going to get a really expensive robot—
Yeah.
—and it’s going to be less safe, it’s going to be harder to manufacture, and then you’re going to have a $10,000 to $20,000 robot there—
Right.
—and you’re going to have a really expensive robot. Let’s call it $50,000. Cost is really a function of manufacturing volumes.
Mm-hmm.
So you really want to build like the car.
Well, that’s why going after the industrial use case is such a no-brainer—
Or even just a home.
Because—
The home needs every—
Well, the home is huge in the end, but if you’re running 3 to 5 times faster than what we’re seeing right now in the home and you kick the cat or something like that, that’s not great. In the industrial use case, everything is taped off.
10. The Billion Robot Economy
Peter H. Diamandis
I remember I was interviewing you for my next book, which comes out in April. Here it is, Where Is God?
Oh, wow.
Peter H. Diamandis
We’ve talked about this, but I’m super excited about it. Of course, you and Figure are prominent in the book.
Hmm.
Peter H. Diamandis
Because this is godlike. I mean, it’s extraordinary. We’re giving life to new systems. I was interviewing you about how many robots there will be and what the price point is.
Yeah.
Peter H. Diamandis
And I want to just double down on that because the numbers are pretty staggering, and they make sense. So if you’re actually getting the price down to $20,000 a robot—I haven’t heard $10,000 a robot, but $20,000 a robot—you’re leasing a robot for $300 a month, $10 a day, $0.40 an hour. And then the question, when you ask the question, is: If it’s really $10 a day, how many would you own or would you have?
Yep.
Peter H. Diamandis
You end up with a lot of robots. So what’s your estimate on the number of robots on planet Earth in 2035 or 2040? Where do you think that’s going?
I think it’s relatively straightforward to think that every human should have a humanoid to do all your work, and then we should have maybe on the order of 5 to 7, maybe 10 billion in the commercial workforce. I think if all goes well, you could basically build tens of billions of humanoids on the planet.
Mm-hmm.
You’re basically building a replica of a human that’s really cheap and works 24/7.
Yeah.
We will be at a point, I hope, in 24 months, where all the robots will build all the robots.
Well, that’s what I wanted to ask about scale. You said we’re going to ramp up to millions a year. One per person on the planet is 8 billion, so millions per year really isn’t that much.
It’s nothing, yeah.
So then you’re like, okay, the self-improvement loop is going to be incredible here.
You also need tons of working capital. If you put a billion robots on the planet, even if they’re, let’s call it, $20,000 apiece, you’re talking $20 trillion of working capital.
Peter H. Diamandis
I mean, you’re not that much—
So there’s some land there.
Peter H. Diamandis
There’s a billion cars on the planet right now. There’s not more than that.
But if you tried to build them in 5 years—
Peter H. Diamandis
There’s a—
It took 80 years to accumulate those cars. Some of those cars are 30 or 40 years old and still running.
We have a couple billion cars on the planet, but we make a billion or more cell phones a year. So—
Yeah.
I think it’s just more cell-phone-like, where it’s going to be personal. We even go back and forth on whether, if your robot breaks, you want a brand-new refurbished robot, or whether you want the old robot you used to have because you’ve known it, you understand it, and it has a personality.
Yeah.
I think it’s going to be with you. It’s going to know everything about you. You’re going to talk to it every day.
Peter H. Diamandis
Yeah. Well, why wouldn’t you just have a personality transfer?
You could, but I think there are some inner workings. It’s got all the scratches on it that you know. It’s your thing, and it’s got a little bit of a feeling. But yeah, for sure, I think that’ll be fine. But at some point—
Wait, wait. Let me ask the geeky finance question, just before we lose the topic here.
Yeah, sure.
So if you have an all-neural-network-based system, it can learn at an incredible rate. The technology is advancing remarkably. You look 24 months in the future, and the demand is on the order of billions, not millions. Like you said, to build that out in 1 iteration—you used the cell phone as an analogy, but Apple had 15 years to profitably ramp up production to 1 billion units a year.
It will be faster. Yeah.
And so the demand is there to do it in 1 year.
Yeah.
But you would need $1 trillion—some insane amount of capital.
Yeah, you’d need a lot of capital.
Peter H. Diamandis
But that’s no longer an insane amount of capital. I mean, we’re seeing—
I think you can—
So what do you do?
I think you can—
Do you leave the world starved, asking for the robot for 5 years—
Peter H. Diamandis
Yeah.
—or do you raise $1 trillion?
If you look at credit card receivables or car leasing, these are trillion-dollar markets per year in terms of financing. So I think the financing market is there for this. What do you do? I think one is, you have to solve the neural-net game. You have to be able to scale with neural nets, and you have to solve pre-training and generalization.
Yeah.
So you have to solve for a general-purpose robot. That is table stakes. You have to solve this. That’s why we’re so obsessed with trying to solve it here at Figure. If you don’t solve that—
Right.
—none of this matters.
Yeah.
The second step is you have to have robots in the loop building other robots.
Mm-hmm.
Those 2 things have to be solved, and you have to design the robot to make sure it can hopefully design itself at the end of the day. There’s a bunch of stuff we’re putting in place in terms of manufacturing execution software, the lines, and all the design of it, so we can, at scale, have humanoids go in, build other humanoids, and get them off the line.
Yeah.
I think these adoption curves are shortening and shortening.
Mm-hmm.
And I do think if we could solve a general-purpose humanoid robot today that could do everything you wanted, I think we could ship 1 billion of them today.
Yeah.
Peter H. Diamandis
What did you say again?
I think we ship 1 billion today.
Yeah, I totally agree.
So basically, it comes down to—
Totally agree.
Can you get the neural nets to work at scale? Can you get the models good enough to generalize to this scale? Build a general-purpose robot—call it a general-purpose robot, like a human in a suit. And then can you get robots in the loop building other robots?
Well, the other thing that’s really compelling is that the neural net is the only IP you need to protect. So as long as you have the federated learning coming back to the mothership and all the training is happening centrally—
Peter H. Diamandis
Yeah.
Do you know the Star Trek Genesis project?
Peter H. Diamandis
Yeah.
You’ve got the little capsule. It has basically the germ of—
Peter H. Diamandis
The DNA.
You could ship literally a box to Kenya that’s like, “Here, here’s the Figure box.” It opens up, and it starts making a Figure manufacturing plant—
Peter H. Diamandis
Sure.
right out of thin air—in the middle of Kenya.
Peter H. Diamandis
Yeah.
And if there's capital there to bring the resources to it, then that's how you get infinite scale.
Peter H. Diamandis
Well, we talk about that. The innermost loop is energy and AI intelligence.
And local mining for the materials or whatever, but it's completely self-contained. But the key—
Peter H. Diamandis
Mm.
—is that you just unlocked that capital that wanted to build something productive—
Peter H. Diamandis
Yeah.
—while all the IP is still flowing back—
Peter H. Diamandis
And 100X-ing the GDP of Kenya.
—to train the neural net centrally.
Peter H. Diamandis
And just drop it.
Yeah, 100X the GDP of that jurisdiction. There's latent capital all over the world.
Peter H. Diamandis
So we talk about how there's a lot of fear out there in the world about losing jobs to AI and robots. The reality is, the conversation has shifted now to, “Well, no, this is going to create massive abundance and universal high income.”
And that happens if, in fact, rather than the company hiring a robot to replace me, I hire a robot to go out and do my work for me—and, in fact, it's able to get triple my salary because it's working 3 shifts, and then earns enough to get a second robot working for me.
Mm-hmm.
Peter H. Diamandis
And so the question becomes: Where is that capital captured? Is it inside the hyperscalers? Is it inside the individual? That's going to be the interesting conversation coming up. How do you think about that, Brett?
Mm.
We're going to sell robots at scale. You're going to be able to deploy as many robots as you want and do whatever you want to do. You can always do whatever you want. No instruction manual: “What do you want it to do?”
It'll learn it. It'll research the internet. It'll use digital tools if it needs to. It'll talk to you. It'll reason. The future's going to be really fun.
11. Safety Unlocks Home Robots
Peter H. Diamandis
Safety and privacy. Let's talk about safety in the home and privacy in the home. There were lawsuits over the last years with Google and Amazon: It's listening to you in your bedroom and so forth.
How do you address safety and privacy? Or is it just going to happen? Is it just too early because we're not there yet?
I think those are some really hard questions to answer in one go, because there's a bunch of different safety implications here that are just—safety's probably the number 1 thing to tackle to get robots into the home at scale.
Yeah.
There's a semantic understanding of safety. If there's a candle lit and I knock it over by accident, or if there's a boiling pot of water and I hit it, just understanding how to be safe in an environment where humans are.
Yeah.
And there's actually the intrinsic safety of: Can the robot be with humans and animals and pets and be safe? That has to be solved. We can talk at length about how we're going to solve those problems.
And then you have the whole privacy, cybersecurity, and other aspects of this that need to be approached with good intentions: How do we solve those problems? We are working on all of those now. They are very difficult things to get right.
Mm.
I do see a path where we can build intrinsically really safe robots around people and pets. We have a plan for how we're going to do that.
Peter H. Diamandis
They could be safer than humans by a large margin, just like autonomous cars are safer than humans at the end of the day.
Yeah. These have superhuman perception. We can see basically all around us at all times. We're always on. We're always computing what to go do.
Mm.
Assuming nobody's trying to be mean to the robots or things like that, I think we should be extremely safe around everything we're doing.
And around privacy, these are going to be in your home, so being upfront about what data we're collecting, where that data's going, and how we're keeping that data private and encrypting that data—all of this is super important.
Yeah.
We have an entire team on cybersecurity here in-house, on both the product and commercial side and the corporate side, working through how we think about this at scale.
Yeah.
Right now, they're great. They're from the big companies that have been doing this for a long time. We think about it as the corporate side as well as the product side, on the robot side as well.
Yeah.
Peter H. Diamandis
Your facility here, which is your sort of prototype manufacturing facility—50,000 robots a year, you imagine?
That facility can support about 4 lines. Each line can do about 12,000 units a year, so a little under 50,000 units a year at full run.
Peter H. Diamandis
What's your next step up, do you think?
We're building thousands of robots right now. That's the big push we're doing right now. You just saw it today.
Yeah.
That's the Figure 02 stuff we're doing off the lines today. Then we want to go to tens of thousands, and then hundreds of thousands and millions. I think we need to take those steps as a company to go do that.
This facility will top out at 50,000 a year, a little under 50,000 units a year at full capacity. Think about the long term: It'd probably be low volume when we look back in 5 or 10 years.
Do you think you might franchise out the neural net and the circuitry around it? All these other people are saying, “I'm building a robot that cleans industrial pipes,” and all these different form factors.
No.
No? Just keeping—
I think it's super unsafe.
Yeah.
I think we see these robots out there like this. They're around humans, and we don't own the hardware. We don't know what they're doing. It's our neural net in it. I think it's—
Oh, interesting.
Yeah. I think it's similar to Archer. When we were building Archer out, I think it's a safety-critical system.
Yeah.
Especially Archer—since they're licensing it out to their folks and stuff like that, that's very problematic.
Yeah.
I think here it's the same thing. Even with humanoids done right, we have a fiduciary duty to our civilization to build really safe humanoid robots at scale.
Yeah.
Giving this AI system or even hardware to anybody that would want it is not something we will entertain.
So then when do you branch out into other form factors, like things that work underwater, things that work—
I think in the future, everything that'll move will be a robot.
Peter H. Diamandis
Mm-hmm.
Mm.
Besides humans.
Yeah.
Within that, I think humanoids will dominate the plurality of all robots. It'll just be such a big percentage of them that the other robots will be niche and expensive, like super-duty trucks that you have out mining. They'll just be made for specific areas, maybe underwater, as you said, or something else.
Or heart surgery or brain surgery. You've got these very, very fine-tuned systems. It's like a robot controlling a robot.
I think you're left with very expensive equipment that's very siloed. You really want to build a general-purpose machine that can learn across a variety of different tasks and have that transfer learning.
Mm-hmm.
I think that's extremely important here, and that needs a very high variety of rich data. This is only going to help the robot system get smarter and better.
Yeah.
So my view is, I think it'll just be humanoid robots everywhere on the planet. There will be other robots there, but they'll just be niche businesses.
Mm-hmm.
Peter H. Diamandis
When I was flying up here, I posted your video that you released on Helix Studio today, and we asked the community for questions. It just blew up with a whole bunch of amazing questions.
One of the questions is: “Do you have a blooper reel, and can folks see it?”
No.
Peter H. Diamandis
“And then what's the weirdest task someone on your team has tried to teach it to do, and it absolutely did not work?” That's from a listener.
Ben Casper, nice. The weirdest task that did not work? Well, the weirdest task…
Peter H. Diamandis
Yeah.
Jogging was fun. Jogging was cool because we really had a steerable jogger, and a lot of this work in running has, again, been open-loop.
Peter H. Diamandis
Mm-hmm.
Actually, I have a gift for you. It goes there.
Peter H. Diamandis
Okay.
I have 2 Figure-deadmau5 hats.
Peter H. Diamandis
Okay. What does that mean?
We basically had robots on stage at a deadmau5 concert at Red Rocks late last year. We generally don’t venture out into weirder stuff. There you go. Nice.
Peter H. Diamandis
Yeah.
We actually had deadmau5 at our last 2 holiday parties at Figure, which was fun. We generally focus on how to design something really useful, but we’ve had some pockets of time to do fun stuff like this.
Peter H. Diamandis
Oh, that’s good.
I think having robots on stage at deadmau5 at Red Rocks was just—I flew in for it. It was unbelievable.
Peter H. Diamandis
Yeah, and you had them on stage for that.
We had them on stage.
Peter H. Diamandis
Oh, that’s great.
It was great. We had several Figure 02s on stage just jamming. We had them all synced, so they synced to the music as they danced, which was really cool.
Peter H. Diamandis
Yeah. I would love that.
They were moving toward it, which was actually wild.
Peter H. Diamandis
I had you on stage last year at the Abundance Summit, but Figure wasn’t with you, so I need to get you back there with Figure in the loop.
Totally.
Peter H. Diamandis
Yeah, for sure. When are we going to see the first Figure in a customer’s home? That’s the next question.
We want to ship robots when they’re really ready. I don’t want to ship slop.
Peter H. Diamandis
Best guess.
I think last year I said 2025. In 2026, we launch a robot to do end-to-end housework in alpha testing, in my home.
Peter H. Diamandis
Mm.
To do full—
Peter H. Diamandis
Like mopping, cleaning, dishwashing—
We’ve done pockets of work really well. We’ve done dishes and laundry and all this. You’re seeing something that’s getting tied together now, but I want to do it across days and weeks of work. I want to be able to drop it into somebody’s home that it’s never seen and also make that work really well.
Peter H. Diamandis
Yeah.
I want to be able to talk to it, and I want it to—
Peter H. Diamandis
Yeah.
—be able to understand me, remember things, show us things, and be able to walk through a room and show it around, almost like a visitor you have at your house for a week, and understand what to go do.
Peter H. Diamandis
2027, 2028, 2029? Your best guess—earliest or latest window?
I’ll tell you what we’re doing: We’re working until midnight every night to solve this problem.
Peter H. Diamandis
Yeah. You were here too.
We are here every weekend, every night, to try to figure out how to solve this.
Peter H. Diamandis
That’s a given. That’s—
This is the question. We call it general robotics. This is kind of where we want to head. I think by the end of the year, we will be able to put a robot into an unseen home and have it do fairly long-horizon work. Then you want to measure how many human interventions you have.
Peter H. Diamandis
Yeah.
Is it once an hour, once a day, once a week, once a month? I think we’ll do that. I think that would be a huge accomplishment for us. I think we’d be on the path to solving general robotics, and then I think next year you’d be on a path where you could ship them into users’ homes and start making sure they work well.
I think anybody who tells you, “Hey, we’re going to ship them or teleoperate them in the home, or we’re going to ship them in at scale in a year,”—you’ve got to ship in a small quantity, and they’ve got to work well. Then you’ve got to work out the problems, and you’ve got to ship again. You have to have an iterative design roadmap, which we have here, and we need to learn.
Peter H. Diamandis
You can learn.
And learn. It’s going to work well at 1. Then it’s going to work well at 10 homes. Then it’s going to work well at 100. It’s going to work well at 1,000. Then it’s going to be 10,000. Then it’s going to be 100,000. Then it’ll be at 10 million. I think it’s going to be super steep.
Peter H. Diamandis
Exponential growth curve.
Exponential growth curve. So I think it’s going to be super steep.
Is there anything to worry about there in terms of time to market? Because the industrial use—you’re sold out for years to come anyway.
Peter H. Diamandis
The question is, is competition going to come in and grab the market before?
Yeah. Or bond with your kids or something.
We feel the work we show today and the work we showed 2 years ago has never been done, in my mind, by any other humanoid company in history.
Yeah.
If that’s the marker, the marker is whenever somebody can do the Keurig test for a couple of minutes with uncut film, and I can watch it do it closed-loop with bimanual manipulation—not even just standing. You’re 2 years away from where we’re at.
Mm-hmm.
So I think we’ll see. We’re trying to push and continue to pull ahead, but hopefully by next year—
By next year—
—we can really show real general purpose inside the robot. Maybe even as soon as this year. I mean, it could happen in a couple of months. We have the right stack now. We’re building datasets at scale so quickly. We’re spending so much time and money on this internally.
We just launched our new B200 cluster. NVIDIA helped; Jensen helped. That went live this year.
How many GPUs are in your—
We have 3,000 B200s that are going live, and we have another set of much larger GPUs that we plan to put out here.
Peter H. Diamandis
Is it for pre-training or—
We just use it for pre-training.
Peter H. Diamandis
—or for inference?
Pre-training.
You do it physically here?
No, we do not use it physically here.
That’s a lot of power.
Yeah, a lot of power.
Peter H. Diamandis
Jay Crate asks a question to the science-fiction geeks among us: What’s beyond Asimov’s 3 laws for you? Have you thought about that? Have you thought about fundamental laws to program into—
Yeah.
Peter H. Diamandis
—your robots?
I think you really want to put these rules down into the nonvolatile memory onboard the robot, at the chip level.
Oh, yeah.
At the chip level.
R2-D2.
Peter H. Diamandis
Mm.
Yeah. We’ve been thinking about this quite a lot. On the one hand, we still want to solve general-purpose-ness. On the other hand, we also want to figure out, once we’re close there, how we get all the supporting things ready to go.
Peter H. Diamandis
Yeah.
Safety needs to be there. Privacy needs to be there. Fleet operations, the reliability of the robot, the maintenance plan for how we’re going to service this, and everything in the business model—all of it, including financing—needs to be packaged and ready to go. So we’re working through all of these now.
Asimov got a lot of things right, and I feel like a lot of the 3 laws—these foundational rules for how we treat humans—we have our own spin on this that I won’t publicly tell today. The goal is to do good work.
Is this something everyone learns internally in corporate training and memorizes and all that?
It’s something that we put, and are going to continue to put, on all the robots.
Peter H. Diamandis
So you have a newborn relative.
A new what?
Peter H. Diamandis
You have a newborn child.
Oh, yeah.
Peter H. Diamandis
Right?
Yeah.
Peter H. Diamandis
So the question here from KK says, “When would you trust Figure to hold your newborn?”
Yeah.
The new Figure 03 is soft. It looks like it’s designed for the home.
That’s this.
But it’s still about 100 pounds, right?
I think it’s the same. I like this question a lot because at Archer I always say, “Until I put myself, my kids, and my family on board, it’s not safe enough to fly anybody.”
Yeah.
I wouldn’t do that today at Archer, and I hope soon I can do that. At Figure, I think it’s the same question: when do I feel safe enough to have a robot in my home?
Peter H. Diamandis
Well, you had it in your home, but—
But I’ve been there. We’ve had folks there, and we monitor it.
Peter H. Diamandis
Yeah.
I think we’re truly safe, but we’re not there now, and I think that’s a great bar for us to hit. When I can put a robot in my home fully autonomously, end to end, around all my kids, I think that’s the point where I would trust it. I think that’s the point where I would say, “This is ready for everybody.”
It’s a good heuristic for us to really try to hit. That’s our goal here: to be able to put it in my home with free rein to go do what it needs to do.
Peter H. Diamandis
Yeah.
We’re there with it now. We babysit it, watch it, and it works well. I’ve shown videos of the robot being with kids, but I think we’re doing it in a safe way.
Peter H. Diamandis
Yeah.
The robots have been totally safe, which is great. Number 1, we need to build a system safety architecture that’s really, really fault-tolerant and redundant in real time.
Peter H. Diamandis
Mm-hmm.
We’ve done that, and we’re doing a better job of that in the future. Number 2, you have to build a safety track record for this. There’s nothing better than actually proving this thing is going to be safe.
It’s a nice barrier to entry, too, if you take the Apple road: it’s got to be a great out-of-the-box experience. That means not stepping on the cat, certainly not dropping the baby, and getting the cybersecurity side of it right, too—not transmitting everything back and having it posted on the internet.
Yeah.
But if you get that reputation—which it sounds like, of all the companies I’ve met, you’re perfectly positioned to get that reputation—don’t make a mistake along the way. Then everybody just says, “You know what? I’m going to choose a Figure robot because I just feel…”
Yeah.
It’s the same way people feel about the Apple brand with cybersecurity.
Yeah.
So—
I hope people walk away from this knowing that general-purpose robots are coming. It feels very close. There are a lot of other things around that you have to get right to build this at scale really well.
Peter H. Diamandis
Is that your main message you want to get across here to everybody watching?
I think the main message we feel every day, if people are excited about AI and robotics, is that this is going to happen really soon.
Peter H. Diamandis
Yeah.
And it’s happening. I mean, you saw it today—
Peter H. Diamandis
But I don’t think people have a clue how fast this transition is going to be.
Just go to our YouTube and watch our videos from the last 2 years. Watch them side by side. It’s dramatic, the change every single year.
Peter H. Diamandis
Yeah.
You saw it today in person. Our robots have now been in customer sites and things. They’ve been out, and we’re going to continue to show more, but it’s hard to feel that because you don’t see it every day.
At some point, probably in San Francisco first, you’ll see more humanoids than humans. I think that’ll be an amazing day.
Peter H. Diamandis
Yeah.
Right now I’m driving in Santa Monica. I was just—
Peter H. Diamandis
By the way, we just did a podcast earlier this morning with Cathie Wood, who sends her best.
Oh, cool.
Peter H. Diamandis
She’s a—
I know Cathie.
Peter H. Diamandis
She’s a huge, huge fan of yours.
Cathie invested in me at both Archer and Figure. She’s great.
Peter H. Diamandis
Wow.
Yeah. She feels the same way.
Yeah, sure.
Peter H. Diamandis
She does. She’s very proud to be an investor in Figure. I was telling her that when I’m out with my kids in Santa Monica, we count the number of Waymos that we see.
Isn’t it crazy?
Peter H. Diamandis
We’ll see 10 Waymos.
Yeah.
Peter H. Diamandis
And then the Coco robots, the little ground robots—
Yeah.
Mm-hmm.
Peter H. Diamandis
The Starship bots and such—I mean, they’re all over the place.
It’s crazy.
Peter H. Diamandis
The first time you see one, you pull out your phone and take a photo. It’s really cool, and then you take it for granted, and then it’s in your way.
Yeah.
Peter H. Diamandis
Right?
My wife and I took a Waymo downtown last weekend for date night, and it was just unbelievable.
Peter H. Diamandis
Yeah.
As an engineer working on these hard projects, I feel like the amount of engineering work they had to do to put it together safely is—I mean, you know what I mean?
Peter H. Diamandis
Google did this, actually.
You control the music and the lights and the environment. If you take a New York City cab and get in the back, it’s like this smoky hell. Then you get into a Waymo, use the app, and turn it into your little paradise. It’s like night and day.
Peter H. Diamandis
Google did such a beautiful job taking the product. Larry Page saw the product win the DARPA Grand Challenge back in 2005, committed to it, and brought the team on.
I think it’s been 16 or 17 years.
Peter H. Diamandis
Yeah.
Yeah.
Peter H. Diamandis
They just stuck with it. Astro Teller at X basically built it out, and then Waymo became an amazing product. It works.
They’ve been undeterred for 16 or 17 years. “Don’t worry about it. We’re just going to make it.” And they did it. It’s unbelievable.
Peter H. Diamandis
Yeah. Amazing.
It’s very inspirational.
Peter H. Diamandis
Kudos to them.
Can I ask you my geeky, sci-fi-meets-geopolitics question du jour? I just got back from Davos on Friday. Today’s Tuesday, so I’m 9 time zones away.
The big topic at Davos, of course, is Greenland, and all the Europeans are saying Greenland could never possibly be mined. It’s impossible to extract minerals from this frozen, cold tundra.
We have some family mining operations in Minnesota, where it’s not nearly as cold but still pretty damn cold.
Peter H. Diamandis
You don’t have mile-thick ice sheets.
We do not have mile-thick ice sheets, but I think if you’re talking about 1 billion and then 8 billion robots, and you need the materials, and that’s the only constraint—
Yeah.
And you have robots that can operate machinery—
Peter H. Diamandis
We can mine asteroids, buddy. Seriously.
You think we’re going to be doing asteroids before Greenland?
Peter H. Diamandis
No. We’ll do Greenland first.
You think we’ll do—
Peter H. Diamandis
Then we’ll do asteroids.
But you think Greenland is viable? I’m not talking about 20 years from now, either. I’m talking about if you want to build 1 billion robots in, say, 6 years from today. That’s realistic.
Peter H. Diamandis
It’s a $50 trillion market for this.
Yeah.
Peter H. Diamandis
That demand drives—
Don’t you think you’d find a way to get through the ice, given 1 million robots working on it?
I’d hope so. Yeah. I think we’d find maybe better physics, but definitely better engineering solutions for this.
Mm-hmm.
Then we’d be able to put an unlimited amount of human capacity at it.
Yeah.
Through humanoids.
Yeah.
Yeah.
That's what I'm thinking, too.
Yeah.
Because the machinery I see is massively automated. It's still driven by people. It's still operated by people.
Yeah.
It doesn't need to be.
It's crazy this shit works, right?
Yeah.
The humanoid neural nets are just—
It's—
It looks—it's just crazy.
The thing is, when you make it work on unloading the dishwasher—
Yeah.
People don't realize how close that is to working on every other task. If you do it—
The dishwasher and folding laundry—these things that we're already doing—are so hard. You have compliant materials that are all changing with you dynamically. Everything isn't in the same place.
Yeah.
It's very different than being on a conveyor system or manufacturing—
Yeah.
Something like that. They already can do it today.
Yeah.
We can do it, and now it's a matter of doing it better—
Yeah.
—doing it with higher reliability across a more diverse distribution of what humans do every day. That's a data play.
The thing is, if you achieve that goal by hacking together 100,000 lines of C++ and teleoperating it, it would look the same, but it would be nowhere near—
As—
—conquering every other problem—
—as flexible, yeah.
But if you did it purely—it's nothing but a neural net—
Yeah.
—and it's purely trained, that means you're within a millimeter—
Yeah.
—of every task—
We are—
—we could possibly define.
We feel like the limiter here is just data. The only difference between it doing logistics and learning towel folding, dishes, or whatever we end up showing in manufacturing is literally just data.
Yeah.
Data goes into the neural net, and now it can do this work.
Yeah.
It's just new neural net weights onboard.
Yep.
I think we're just bound by data now. I think it's not a trivial thing to do to get the right pretraining set for this at scale, but we have a bet that I think will work. We've been deploying that at scale for the last 3 or 4 months.
Yeah.
Stay tuned. We're working through it. I hope this will lead to—I think you'll see a lot of positive transfer emerge from a robot that's able to generalize to a lot of things.
Yeah.
Peter H. Diamandis
Yeah. Amazing. One last thing before we wrap up. Can we pull the camera in close and maybe give us a tour of Figure 03?
Yeah, let's do it.
Peter H. Diamandis
Thanks for the close-up and intimate tour. So, Figure 01.
Figure 01. One cool thing about Figure 01 is that we designed most of the system in-house. We didn't care about looks. We cared about unlocking the AI and controls team. It was something they could use from a software perspective.
Peter H. Diamandis
Yeah.
We designed and walked this robot in under 1 year, so I incorporated the company. We think it's probably one of the fastest times in history.
That's a lot of parts, man.
Peter H. Diamandis
Did you draw this by hand early on?
These are internally made, actually. David, our design lead, designed this. It's not his prettiest robot, but I think it had what we needed, which is a functional robot we could get off the ground and start using for all the AI policy development. We did the Keurig K-Cup with this robot.
Peter H. Diamandis
No way.
Can I move the hand, too?
You can definitely move it, yeah.
Peter H. Diamandis
Are you sure? This is going to be a collector's item someday. You break this thing, you bought it.
Yeah. It's heavy. It's about 130 or 140 pounds.
That's not that different from—
Yeah, not bad.
That's all aluminum?
It's all aluminum.
Peter H. Diamandis
All CMC.
We CNC aluminum. Most of the structure is done.
Peter H. Diamandis
Yeah.
Mm-hmm.
Yeah.
Peter H. Diamandis
What else should we know about this before we move to Figure 02?
We cared about speed, so we didn't really care about the wiring and some electronics. A lot of the design was mostly just to get a functional humanoid robot out so we could do development on it.
Peter H. Diamandis
Right.
We did that. We built a few of them. We did our first neural network on this robot, which I think was phenomenal. We did so much development with it really quickly.
Peter H. Diamandis
Yeah.
We also learned how to build actuators, battery systems, wiring, structures, kinematics, joints, and different sensors. All of this is stuff we learned. Then we used it all and integrated it into Figure 02.
But you got the cost down from Figure 02 to Figure 03 by 90%. What was the cost from here to there? Probably another 90%.
About the same, to be frank.
Wow.
A lot of it was machine parts, and we moved a lot of the tooling parts to Figure 03.
Peter H. Diamandis
You've got 2 cameras here.
2 cameras here. We have a back camera—
Oh, on the unit?
You see? Yep. We also have cameras right here in the torso pointing down.
Amazing.
So we can see where the feet are in case you have a box occluding them.
Peter H. Diamandis
Come take a look at the camera in the back of the robot here for 1 second.
Where's the camera pointing down?
It's right there in the pelvis. See, right here.
Peter H. Diamandis
Back here, you've got—what's going on here? There are camera ports here?
Yep. We obviously have a backward-facing camera. We have different ports for debugging, if we need to hook up a cable to it, and we can also turn the robot on and off from here.
Amazing.
Basically, we moved all the wires internally into this robot. All the structure is an exoskeleton, so all the exterior loads—almost like my aircraft at Archer—the skin, the outside housing, handles the loads.
Peter H. Diamandis
Sure.
We have the same thing here.
Peter H. Diamandis
Right.
All the outer shell took all the loads. We have our second-generation actuators and third-generation hands on this robot. We have more cameras onboard, about double or triple the amount of compute, and about double the battery capacity onboard.
Peter H. Diamandis
Yes. The degree of beauty went up.
Yeah.
Peter H. Diamandis
Significantly.
Yes. Our design lead did a good job making this much more presentable.
It's funny that it's venting heat out the armpits, just like a real—
Yeah. It actually sucks air in here.
Okay.
We push it out through the torso at the bottom.
Okay.
Yeah.
What's going on in the back?
We basically have different padding on the knees and some parts of the arms to make it so that if you got your finger stuck here—
Oh.
—
Safety.
Yeah. Maybe it would hurt your finger, but it wouldn't cut it off.
Yeah.
You know what I mean? It's similar to what you see on a car door today.
Sure.
Peter H. Diamandis
And here's the workhorse.
This is our Figure 03.
Nice.
We made the robot much skinnier and lower mass, but kept all the speeds and torques the same. It's just as powerful and just as fast, but also skinnier—
What's the mass now?
—and takes up less space. This is about 135 pounds.
135.
This is about 150, a little over 150 pounds.
Yep.
Peter H. Diamandis
Carrying weight—how much weight can it carry?
It's a very different hand.
About 20 kilos.
Peter H. Diamandis
20 kilos.
Yeah. It's a completely different hand. The hands have a glove, tactile sensors, compliant material on it for better grasp, and a camera. All the parts—or most of the robot—are soft-wrapped. You can see it even up here, with a squishiness to the chest and different parts of the robot. We have no, or very few, pinch points in the robot. What else? We reduced the cost massively. We have a better thermal system and compute system. We also increased compute on this robot from the last generation. We have new feet that have a toe.
Peter H. Diamandis
Sure.
You might think the toe is—
Peter H. Diamandis
Yeah, no, it's a major part of the gait.
It's helpful as a passive toe on the foot, but you might think this helps it walk better. It's not just that. When we get down here, we're on our toe box. It really helps us get the range of motion.
Oh, it's nice.
Without that, you might need more joints or something else on the robot.
This is just a totally flat foot over here.
Peter H. Diamandis
Brett, talk about the face, because this is a big question: Do you develop or show facial features or not? And you went—
Yeah. What do you think? What do you have—Westworld, or I, Robot?
Wow. I mean, it comes across as beautiful, right? A high degree of beauty. It comes across as sleek, but it could have a negative, a little dystopian feel with the black face.
Yeah. We have 3 screens on the robot. This is powered off. We have a main screen and 2 screens on the side. Then, obviously, we have a bunch of cameras and sensors in the head. On the screens, we can basically do anything. You could watch a Netflix movie.
Oh, with a brain—
Look into my eyes.
The brain is the screen.
Yeah, exactly. Whatever you want. If kids get bored, it's like, “Let's throw some up there.”
So the brain is right in here, which makes a ton of sense to me. And it's where the ancient Romans thought the brain dominated the heart.
You basically need a lot of onboard computation. There is nowhere else to put it right now.
Yeah, exactly. It's also easier to vent the heat from here, too. Then you just put all the sensors up here, and it makes total sense.
Peter H. Diamandis
I guess I could put a latex face over the head if I wanted.
Yeah, you can basically put a silicone face on it and put hair on it. We could go—
Peter H. Diamandis
I could get to your hair.
Yeah, yeah. We also have other outfits. This is one of our logistics bots. It's basically the same robot, but we're able to outfit it with different types of soft goods. We have another robot here that we've also put to work; it's wearing a jacket. This is cut-resistant. So they all have different—
Oh.
—different traits. Some of these gloves are also better for grasping different materials.
Peter H. Diamandis
Mm-hmm.
Those might be, say, dusty, or maybe a piece of sheet metal, or slick.
Do you think it would operate in zero-G? You just need a better training set and—
I think so. We'd really love to run—
Peter H. Diamandis
I've got a zero-G airplane.
—robots at scale in space.
You're gonna populate the universe.
Peter H. Diamandis
I've got my zero-G airplane. We could—we should—
We should test it.
Peter H. Diamandis
—take it up.
Yeah, let's get these things on there.
Peter H. Diamandis
Yeah.
Oh, that'd be a great test.
Yeah.
Peter H. Diamandis
Yeah.
Well, look, we're going to build data centers in space very soon. Somebody needs to assemble them. Zero-G is the operating—
And then we'll get other planets, too. It'll be super important.
Peter H. Diamandis
Yes. Yes. That's—
A lot of materials.
Peter H. Diamandis
—yeah. And then we'll disassemble the moon and the asteroid belt and use it for materials.
100%.
Peter H. Diamandis
Oh.
Alex will love that you said that.
Let's do it.