Read claim 1, not the keynote. Figure AI's grant US12638859B2, "Bipedal action model for humanoid robot" (issued May 26, 2026; inventors Corey Lynch, Toki Migimatsu, Michael Ahn), is not a patent on a humanoid that walks. Plenty of prior art walks. What the claim language fences off is narrower and more interesting: a model that maps a high-level objective onto the low-level action stream a two-legged machine needs to actually move.
The CPC tells the real story. The classifications attached to this grant put control of position and course (G05D 1/495) alongside legged-locomotion vehicle mechanics (B62D 57/032) — and, tellingly, a natural-language element (G06F 40/40) and visual-recognition training classes (G06V 10/766, G06V 10/774). Strip the marketing and you get the actual invention: the bridge between a language-shaped instruction and a bipedal gait. That bridge — not the legs — is what Figure chose to protect.
Why does the distinction matter? Because a title fences off nothing. "Bipedal action model" sounds like it might cover any humanoid that takes instructions, and a casual reader could call this a blocking patent on instructable robots. The independent claim won't support that. It supports the specific architecture the inventors described: an action model conditioned on a goal, emitting the control signals for legged locomotion. Competitors who reach instructable bipedal control by a materially different route are not obviously inside this fence.
This is also a signal about where Figure believes its defensible IP lives. A humanoid company can file on actuators, on hands, on power systems — the mechanical commodities everyone needs. Figure filed here, on the software layer that turns intent into gait. That's a bet that the differentiator in humanoids is increasingly the model, not the metal, and it lines up with the broader sector move from hardware demos toward learned control.
The honest caveats, in this desk's house style. First, this is a granted patent, not a published application — the claims survived examination, which makes the scope more settled than a pending case. Second, a claim describes a method, not a shipped capability; nothing here proves the gait works outside a controlled setting. Third, the language element means prior art in language-conditioned policy learning is exactly where a challenger would dig. But the filing choice itself is the tell: Figure is fencing the language-to-locomotion layer, and that's the layer worth watching.
For anyone tracking who owns what in humanoids, the move is to verify against the record rather than the reveal video. The grant is right there on PatentBear; read the independent claim, check the CPC, and ask what's actually enclosed. Here, it's the action model — the part that decides how a bipedal robot moves in response to a goal.