The hard claim in humanoid robotics is not a leg or a hand in isolation — it is the system that coordinates dozens of joints in real time from what the robot sees and is told to do. So when a humanoid company is issued an enforceable claim that ties a learned control policy to a specific, counted joint architecture, that is a defensible position worth reading carefully. On May 26, 2026, Figure AI was issued US12638859B2, covering a bipedal action model for a humanoid robot.

The claim is unusually specific about the body it controls. It describes a humanoid with a torso, two arms and two legs providing at least 30 degrees of freedom, a sensor suite of cameras and proprioceptive joint encoders, and a two-tier control stack: a higher-level model that ingests multimodal input and emits a token sequence describing task intent, and a lower-level model that conditions on those tokens and the robot's current pose to output continuous action chunks, which a controller converts into actuator signals.

The present disclosure provides a humanoid robot system comprising a mechanical structure including a torso, two arms, and two legs providing at least 30 degrees of freedom, actuators coupled to the degrees of freedom, a sensor suite comprising at least one camera and proprioceptive sensors including joint encoders and an inertial measurement unit— Bipedal action model for humanoid robot, US12638859B2

The classification reflects the hybrid: G05D 1/495 and G05D 2101/15 mark autonomous-machine control, B62D 57/032 marks legged locomotion, and G06F 40/40 marks the language-model component — a claim written at the seam between a learned policy and a walking machine. For a reader mapping where humanoid content is being locked up, this is coverage on the brain-to-body link, not just a body part. The "at least 30 degrees of freedom" figure in the claim is not decoration; it is the kind of concrete limitation that gives a patent its shape. Degrees of freedom — the count of independently controllable joints — is the rough currency of humanoid dexterity, and tying a control-policy claim to a specific minimum count anchors the coverage to a class of machine rather than to a vague idea of a robot. The two-tier structure the claim describes, a higher-level model emitting task-intent tokens and a lower-level model emitting continuous action chunks conditioned on current pose, mirrors the architecture pattern that has become common across embodied-AI research. Holding an issued claim that pairs that pattern with a counted-DOF humanoid body is what makes the grant a fixed point rather than an abstraction.

The portfolio underneath the policy

The action-model grant does not stand alone; Figure's issued record fills in the body the policy commands. A June 9 grant, US12649246B1, claims a humanoid ankle region with separate foot-roll and foot-flexion actuator assemblies controlling pitch and roll of each foot — the joint mechanics that a bipedal policy ultimately has to drive. US12611766B2 covers advanced kinematics, including a hip-flex actuator geometry and a design that omits a distinct torso-pitch actuator, using the hip actuators to provide torso pitch instead — a specific architectural choice now held as issued coverage.

The hands and upper body are in the record too. US12605824B2 claims a humanoid with a multi-degree-of-freedom finger-and-thumb end effector and a wrist that rotates across a wide angular range, and US12611767B2 covers an efficient inverse-kinematics control method that uses a "phantom" task to keep the robot within safety-related limits while performing its real task. An earlier related grant, US12578733B2, also claims a two-model (alpha/beta) bipedal action architecture, indicating the control-policy idea is held across more than one issued patent.

What the spread of classes shows

Read together, Figure's issued footprint sorts into a coherent shape: the learned control policy (the bipedal action models), the kinematic architecture (hips, the torso-pitch choice), the actuated joints (ankles), the end effectors (hands and wrists), and the supporting structure (head-and-neck assemblies such as US12539618B1). The portfolio's classification facets cluster heavily in the B25J manipulator classes — B25J 9/0009 and B25J 11/0015 recur across the grants — alongside the B62D 57/032 legged-locomotion class. A company holding issued claims from the action model down to the foot actuator is the data point.

That coverage is recent and concentrated: Figure's grants in the patent record cluster in the last two calendar years, consistent with a young company converting a fast filing program into issued patents. The bipedal-action-model grant is the keystone of the set — the patent that connects the rest, claiming the policy that turns intent into the actuator commands the other patents' joints execute. One structural detail in the kinematics grant is worth surfacing because it shows how specific this coverage gets. US12611766B2 describes a humanoid that omits a distinct torso-pitch actuator and instead achieves torso pitch through the hip-flex actuators — a deliberate design choice that removes a motor and reallocates its function. Whether or not that choice proves optimal, holding an issued claim on it converts an engineering decision into a fixed position in the record, the kind of granular coverage that distinguishes a portfolio built around a real machine from one built around a concept. Stacked with the ankle, hand, and head grants, it sketches a body whose major subsystems each appear as separate issued claims, all of which the action-model patent's control policy ultimately drives.

None of this speaks to how the claims will be enforced, or whether they will be. What the late-May record establishes is narrower and firmer: as of that date, Figure holds issued coverage spanning the humanoid control policy and the mechanical body it commands, documented in patent numbers with the company on the assignee line.