Figure Says Its Figure 04 Humanoid Hit Design Lock

Figure’s fourth-generation humanoid, the Figure 04, has reached “design lock” with parts already shipping, chief executive Brett Adcock said in a public recording of an internal all-hands meeting published on 22 October. The hour-long session, reported by the trade outlet Humanoids Daily, doubles as a statement of intent: the California company is pouring billions into data and computing to train a robot it has barely shown in public.

Every number that follows is Figure’s own, disclosed by Adcock and not independently verified. Design lock means the hardware specification is frozen, but Adcock said the robot is still moving through component testing — its actuators running in the test lab and ten full battery packs “brought up” the week before, with load and thermal testing to follow.

The pitch is data and compute, not a finished robot

Adcock’s core argument is that general-purpose robots will come from scale, not hand-tuning. Figure says its Index programme — an app that pays people to upload video of everyday tasks — has gathered 37.5 million videos across 112 countries, with an estimated $8.1 million spent on data in a single week. He put the company’s cash balance “above $2 billion” and described a compute deal with Nscale starting at $3.5 billion and intended to reach $6 billion. Figure says it built its 1,000th Figure 03 over the summer and cut costs by more than 90% from the Figure 02, though it gave no absolute price. Index, launched in August with a reported 16 million videos, has since drawn 1.26 million app downloads, and Adcock said 210,000 people uploaded at least one clip in the past week. He also said robots were operating in three US states and that a 200-hour logistics run was kept going by swapping robots in to charge, rather than one machine working 200 hours straight. It is the same race TSN has tracked at Tesla, which says Optimus production has begun, at Unitree, whose H2-X “Guardian” clip was staged, and among startups such as Zurich’s Enact, chasing a sense of touch.

Figure remains private; it raised more than $1 billion in a Series C at a $39 billion post-money valuation in September 2025, and has deployed earlier robots with BMW Manufacturing and the retailer Catalyst Brands. Adcock said the roadmap had “completely changed” over the past year and the team had doubled.

What the all-hands showed, and what it carefully did not

The gap between claim and evidence is the story here. “Design lock” and “parts shipping” are real engineering milestones, but Figure showed the Figure 04 only as a blur and offered no independent dexterity benchmark for its much-hyped new hand, which Adcock said approached “human parity”. The data figures are striking, yet Humanoids Daily noted that uploads are not the same as cleaned, annotated footage usable in training — 37.5 million videos is an input, not a result. Even Figure’s own transfer-learning evidence is mixed: its Helix 2.5 model succeeded in about 56% of trials across trained chores in unfamiliar homes, progress that still leaves a wide reliability gap. The cash balance and the multi-year compute commitment are separate figures, not a freshly raised round of either size. And Adcock’s description of an internal model as “baby AGI” is a founder’s phrase, not a demonstrated capability; the presentation disclosed no evaluation detail to support it. For UK readers weighing the humanoid hype, the signal to watch is not the valuation talk but whether Figure lets outsiders test the Figure 04 and publishes reproducible numbers. The next marker Adcock named is Helix 3.0, teased on screen with no release date.

Alibaba’s Amap is testing a similar idea in public research, with a humanoid model trained partly on 500 hours of first-person human video.