A Humanoid Robot in Every Home? It's Closer Than You Think w/ Brett Adcock (at A360 2025) | EP #156
Summary
Adcock’s core thesis is that humanoids are the “ultimate deployment vector for AGI,” because intelligence trapped on a server must ask humans to affect the physical world. A mechanical-human form factor can operate across human environments, while one foundation model can transfer across tasks without hardware changes. He says solving the key challenges would put Figure in the right decade for the space’s “iPhone moment.”
The near-term commercial thesis is repetitive workforce labor, not the far more complex home. Adcock puts human labor at roughly half of global GDP, which Diamandis translates into a $50-$60 trillion TAM against $110-$120 trillion of global GDP. Adcock says that if Figure had 100,000 functioning robots today, its two first customers would take them, while home operation remains “the Wild West.”
Figure now has robots working daily at BMW’s Spartanburg plant, autonomously placing sheet metal onto fixtures with what Adcock calls “no faults, no failures.” The BMW task took about a year to execute fully end to end at high speeds; Figure 01’s cycle fell from roughly four minutes last summer to 40 seconds. With Helix, Figure completed an unnamed logistics customer’s use case from nothing in under 30 days and believes a repeat could take less than 48 hours.
Helix is the claimed AI inflection, turning robot programming into speech-directed generalization. Trained with about 500 hours of data, two robots used an English instruction to put away groceries deliberately withheld from training. Adcock calls it “probably the most important AI update for robotics in human history,” though the evidence presented is Figure’s own demonstration.
The hardware road map combines rapid iteration with a mass-market cost target. Figure designs a new platform every 12-18 months; Figure 3 took 18 months and is described as “90% cheaper,” smaller and lighter, with improved sensors and neural-net-oriented hands, head and feet. Diamandis raises a future $20,000-$30,000 price point; Adcock says bill-of-materials work does not indicate that the product should be extremely expensive, but does not explicitly reaffirm that range. Diamandis frames a $30,000 robot as roughly $300 monthly, $10 daily or $0.40 hourly when continuously available.
Adcock has moved the home timeline forward by “multiple years,” but makes it conditional on data, generalization and safety. Figure will start internal home alpha testing this year, with robots expected in homes “in the coming years.” The goal is a robot that understands spoken requests and performs hours-long tasks without repeated prompts or fixes—not merely one that manipulates familiar objects.
Execution depends on an unusually intensive, vertically integrated organization built for hardware speed. Adcock initially removed near-term financing risk with his own capital, reached a $1 million monthly burn within six months and recruited around an “iPhone moment” despite telling candidates the success probability was “pretty low.” Figure’s office-based, five-to-seven-day culture is organized around one reward: “We want to ship product”; Adcock says he does not do 1-on-1s.
Deep dive
1. A humanoid gives AGI agency in the physical world
Adcock sees an AGI confined to servers as “a really negative, almost dystopian future”: however intelligent it becomes, it must ask or direct a human whenever it wants something done physically. A humanoid supplies the body—and therefore the physical-world agency.
The human shape is not cosmetic. His requirement is one mechanical platform that, without hardware changes, can operate in human environments and learn many applications through transfer learning, ultimately using one foundation model to control the robot end to end.
Diamandis’s crucial challenge was the probability of success. Adcock’s answer—“pretty low”—rested on three gates: reliable hardware at human speed and range; imitation learning because “this is a neural net problem, not a control problem”; and speech-directed generalization to unseen tasks through one network. Diamandis said those requirements looked “pretty dire” in 2022; Adcock later said Figure had solved—or was making substantial progress on—all of them.
2. Hardware velocity requires owning the entire stack
Diamandis notes that Figure went from a cold start to shipping its first robot in 31 months; Adcock says it was walking within 12 months of filing the corporation. His rule is blunt: “The first or second generation hardware is always going to suck,” so Figure targets a new platform every 12-18 months.
Diamandis frames vertical integration as a necessity because no ready-made humanoid supply chain existed for motors, actuators, batteries, sensors or kinematics. Figure consequently does the hardware, firmware, embedded systems, operating systems, controls, AI, testing, manufacturing, integration and fleet operations itself.
Figure 3 represents another full redesign after 18 months. Adcock describes it as “90% cheaper,” smaller and lighter, with better sensors and hands, head and feet designed for neural nets; manufacturing was scheduled to begin this year.
Adcock funded the opening years himself and hit a $1 million monthly burn within six months. Recruiting offered near-term funding certainty alongside an intentionally demanding culture: office attendance is mandatory, teams work five to seven days, and the shared cultural “dopamine” is shipping. Adcock says he does not do 1-on-1s.
3. Commercial labor is the first market—and demand is not the constraint
At BMW’s Spartanburg, South Carolina plant, Figure robots work daily placing sheet metal onto fixtures. Adcock says the task is fully autonomous, meets the required performance speed and runs with “no human intervention, no faults, no failures”—or, as Diamandis adds, “no days off.”
For the two first commercial customers, Adcock says that if Figure had 100,000 working units today, they would take all 100,000. Beyond them, he says he could sign 50 Fortune 100 customers by the weekend, but cannot supply them.
The economic frame is enormous but explicitly long-term: Adcock calls commercial human labor roughly half of GDP, and Diamandis estimates a $50-$60 trillion addressable market from $110-$120 trillion in global GDP. Workforce jobs also repeat and can support materially higher robot pricing than households.
4. Helix compresses task deployment toward hours
Diamandis frames the pivotal decision as Figure moving from baselining the AI systems of large investor OpenAI to building AI internally. The result, Helix, is a vision-language-action model intended to connect natural-language instructions directly to robot behavior.
In the home demonstration, the instruction was simply to “put the groceries away.” The grocery items had been withheld from training, yet two robots identified where items belonged and coordinated through a single neural network on each robot; Adcock says Helix used only about 500 hours of training data.
Their handover behavior was not manually staged: the robots learned to look at one another at the instant one should release and the other grasp. Adcock calls the gaze a learned clearance signal and argues that nods, gestures and visible attention will be as important as grasping when robots integrate into the world at scale.
Helix also changed Figure’s commercial learning curve. Adcock says BMW took a year to execute fully end to end at high speeds; he also says Figure 01 took four minutes last summer and now takes 40 seconds. The second customer’s task went end to end in under 30 days. Adcock believes Figure could redo it in less than 48 hours and that robots will learn new work “in the matter of hours” this year.
5. The home opportunity arrives only after safety and semantic grounding
Adcock rejects the idea that industrial success transfers automatically into homes. Factory work resembles highway driving; the home resembles city driving, with changing layouts, unfamiliar objects and semantic hazards such as knocking over a candle and burning down the house.
His diagnosis is now “data bound.” As evidence of semantic grounding, they put a moving, singing cactus toy before Helix and asked it to “pick up the desert item”; the model connected cactus, desert and toy despite the odd presentation. Adcock thinks increasing the training data by a couple of orders of magnitude would probably make the home system work.
Figure will start alpha testing in Adcock’s and engineers’ homes this year. His forecast remains “this decade” and “in the coming years”: users should eventually speak to a robot and receive hours of autonomous work without further correction, taking over chores humans currently handle alongside household appliances.