Every humanoid pitch deck sells the robot. The patent Figure AI was granted on July 21, 2026 is not about the robot. US12686136B2, titled “Humanoid robot data collection system,” issued to Figure AI with named inventors Victor Ragusila, Nathan Jenest, Vadim Chernyak, Corey Lynch and Toki Migimatsu, and is classified under B25J 13/025 and B25J 9/1689 — the manipulator-teleoperation and program-control corners of the robotics class. It is the company’s only grant in the July 21 issue, and what it covers is the apparatus a human wears while piloting a humanoid, plus the tie between that apparatus and the training data it produces.

That distinction is worth being precise about, because the record is not internally consistent. The abstract describes a three-part system — “a wearable data collection apparatus, a computer, and a humanoid robot in data communication via a network” — and lists a base mount, articulated arms extending from the base mount, sensored gloves and a piloting control system as the apparatus contents. Claim 1, the operative scope, recites something considerably narrower and differently centered. It is an apparatus claim to a mobile wearable data collection apparatus comprising a headset with a virtual- or augmented-reality display, a glove of flexible textile with haptic feedback, a rigid housing on the dorsal side of that glove, a finger encoder mounted at the housing, and a deformable connector pivotably coupled between the encoder and the fingertip. The networked computer and the robot itself appear in the abstract as background description; they are not elements of claim 1.

The hardware that most readers would picture as the “rig” sits in the dependent claims. Claim 6 adds an articulating arm worn from the operator’s torso to the hand, terminating in a glove mount that detachably couples to the rigid housing. Claim 7 adds the base mount and an articulated arm carried by the operator without external support while walking. Claim 8 breaks that arm into rigid frame links pivotably coupled between base mount and glove mount. Claim 9 adds a head position sensor and a piloting control system carried by the operator that reads finger and head position. The exoskeleton, in other words, is claimed — but as build-out on a glove-and-encoder core, not as the independent claim.

The commercially load-bearing element is claim 10, the last one, which is where the hardware is connected to the model-training pipeline.

The mobile wearable data collection apparatus of claim 1, wherein the mobile wearable data collection apparatus is configured to: further collect robot data generated by the humanoid robot, relate the collected robot data to at least data generated by the finger encoder, and provide the collected robot data for training an artificial intelligence model.— Humanoid robot data collection system, US12686136B2

The collection layer and the model layer share inventors

Read that claim next to the rest of the indexed footprint and the shape of the portfolio changes. Two of the five inventors on the new grant, Corey Lynch and Toki Migimatsu, are also named on US12638859B2 and US12578733B2, both titled “Bipedal action model for humanoid robot.” Those two records are directed at the learned side of the stack: US12638859B2 describes a hierarchical model in which a Beta model turns multimodal input into a token sequence and an Alpha model emits continuous action chunks, while US12578733B2 describes the two models deployed on separate GPUs and jointly trained. Lynch is additionally named on US12605824B2, a wrist-and-end-effector grant covering rotational range at the left wrist and a thumb assembly with at least three degrees of freedom.

So the same small group of names appears on the apparatus that harvests demonstrations, on the end effector those demonstrations are recorded through, and on the models the demonstrations feed. Figure AI has now been granted claims at all three points of that loop rather than only at the robot end of it. For a company whose product is not yet a shipped consumer good, coverage over the data-acquisition step is the part of the estate that touches the training input, which is the scarce resource in embodied AI.

What the wider footprint indexes to

Searching issued US grants by assignee returns 14 records across three name variants — 11 under “FIGURE AI INC.,” two under “FIGURE AI” and one under “Figure AI Inc.” The bulk of that indexed set is mechanical. US12611766B2 covers a leg architecture in which hip flex actuators are angled below the transverse plane and the robot lacks a distinct torso pitch actuator. US12649246B1 covers an ankle region with separate roll and flexion actuator assemblies. US12667974B2 covers a head and neck assembly with frontal and rear shells around an electronics assembly. US12420434B1 covers an end effector built from identical, self-contained finger assemblies each driven by a single electric motor, and a further record covers a phantom inverse-kinematics task combined with a safety-related limit.

Against that backdrop the new grant is a departure in classification as well as subject matter. The most frequently recurring class in the indexed set is B25J 9/0009, on six records, followed by B25J 11/0015 on four and B62D 57/032 — legged locomotion — on three. Neither B25J 13/025 nor B25J 9/1689, the two classes carried by US12686136B2, is among those leading buckets. The cadence is also visible: ten of the 14 indexed grants carry 2026 issue dates, running 2026-02-03, 2026-03-17, 2026-03-24, 2026-04-21, two on 2026-04-28, 2026-05-26, 2026-06-09, 2026-06-30 and 2026-07-21. The remaining four issued across 2025, on 2025-07-22, 2025-09-02, 2025-09-23 and 2025-10-21. More grants have issued to the company in the first seven months of 2026 than across all of 2025.

One caveat on the count. Assignee search matches indexed assignee strings, and a substantial share of US records carry no assignee value at all, so 14 is the number of grants attributable to Figure AI by name in this index, not a complete measure of what the company holds. Published applications, which lag issuance, are not counted here either.

Two things follow from the record as written. First, anyone characterizing US12686136B2 from its abstract will describe a networked robot-and-computer system; the claims describe a wearable apparatus whose independent claim turns on a finger encoder and a deformable fingertip connector, and claim scope is what the claims say. Second, the grant places Figure AI’s issued coverage on the teleoperation-and-capture step, adjacent to the model patents rather than duplicating them. In a field where the mechanical patents describe machines that are still pre-revenue, a claim that reaches the demonstration-gathering apparatus and its link to model training is a different category of asset from a claim to an ankle joint.