Brett Adcock: Humanoid Run on Neural Net, Autonomous Manufacturing, $50T Market #229
Summary
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.”
Deep dive
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.