Figure AI Patent US12578733B2: The Dual-AI Brain Behind Humanoid Robots
💡 Figure AI's US patent US12578733B2, granted March 17, 2026, covers a "Bipedal Action Model" (BAM) - a dual-AI architecture that splits a humanoid robot's brain into a slow-thinking language model and a fast-reacting motor controller, each running on a separate GPU. The architecture gives a robot with 62 joints the ability to understand a spoken instruction and move fluidly in real time - simultaneously. If this pattern becomes the industry standard, it could position Figure AI as the toll-booth operator for an entire wave of general-purpose labour robots now entering factories worldwide.
What the patent actually describes
Humanoid robot patent US12578733B2 covers a control architecture called the Bipedal Action Model (BAM). Rather than relying on one monolithic neural network that tries to do everything at once, BAM splits the robot's cognitive work between two specialised models running on separate graphics processing units. The application was filed on September 4, 2025, published as US20260064125A1 on March 5, 2026, and granted just twelve days later on March 17, 2026 - a fast-track timeline that suggests the claims were narrow enough to sail through examination cleanly.
The invention is assigned to Figure AI Inc., the California-based humanoid robotics startup that valued itself at $39 billion in its September 2025 Series C round. The core claim covers a robot with 62 degrees of freedom across its full body - arms, hands, torso, legs, and head - controlled by a layered AI system that can receive instructions in natural language and produce smooth, continuous joint commands in real time. That combination - language in, fluid motion out - has been the hardest unsolved problem in practical robotics for a decade.
Understanding what the patent claims requires understanding how previous systems failed. That is the most important context.
The problem: why robots have been so clumsy until now
A robot that can understand "pick that box up carefully and place it on the shelf" must do two very different cognitive jobs at the same time. The first is semantic: what does "carefully" mean, where is the box, what counts as "the shelf"? This requires the kind of world knowledge that only large language models have demonstrated at useful scale. The second job is physical: adjust grip force to 4.2 Newtons, rotate wrist 37 degrees, shift weight to left foot - all within milliseconds and continuously as the object moves and the environment shifts.
The tragedy is that language models and motor controllers run at incompatible speeds. A large language model processes a thought roughly once every 50-1,000 milliseconds - barely real time by robotic standards. A proper motor controller must update at 100 Hz or faster to avoid instability. You cannot run one neural network at both speeds simultaneously without severe compromises in either reasoning quality or motion smoothness.
Older systems either used a slow, centralised controller (resulting in jerky, stop-start motion) or a fast, shallow controller that could not understand language at all. The BAM patent formalises an architecture that solves this by making the speeds explicit - and by filing a patent on the solution, Figure AI claims the design as defensible intellectual property.
The alpha-beta architecture: two minds, one body
The patent describes two complementary models. The "alpha" model is large - billions of parameters, trained on internet-scale data plus simulation and teleoperation recordings. It operates at 1-20 Hz, handling language understanding, long-horizon task planning, and semantic reasoning. Think of it as the robot's prefrontal cortex: it knows what the task is and roughly how to accomplish it.
The "beta" model is compact and fast. It receives a compressed representation - a latent vector - of the alpha's semantic intent, and turns it into continuous control commands for all 62 joints, running at 100-10,000 Hz. The beta generates what the patent calls "action chunks": predicted sequences of joint positions, velocities, and torques spanning 50-150 milliseconds into the future. Predicting a short trajectory rather than a single instant dramatically reduces the compounding errors that make imitation-learning robots drift off task over time.
The result is a robot that moves the way a skilled human does: smoothly and continuously, adapting to physical reality in real time, while still being guided by high-level language instructions that arrive much more slowly. This decoupling is the core of the patent's novelty - and the reason competitors will have to find a different path or negotiate a licence.
What this patent depends on - and what it could unlock
The BAM architecture is not self-contained: it sits at the intersection of at least three parallel innovation streams, and its value depends on all three maturing together. First, it requires high-bandwidth GPU compute. The alpha model runs on a dedicated GPU - the same class of silicon that NVIDIA sells for AI inference. It is not a coincidence that NVIDIA is one of Figure AI's Series C investors: as the GPU provider, NVIDIA has a direct commercial interest in this architecture scaling. Second, the patent explicitly covers a layered training framework - combining internet-scale video, physics-simulation data, and real-world teleoperation recordings. This pipeline depends on advances in simulation fidelity, sensor miniaturisation, and data collection infrastructure that are all moving fast in parallel.
Third, the deployment model described in the patent supports three configurations: fully local (both models on the robot), fully remote (both in the cloud), or hybrid. The hybrid mode is the most commercially interesting: a lightweight beta model on the robot, a powerful alpha model in the data centre, communicating over low-latency wireless. This is exactly the kind of workload that 5G and emerging 6G networks are designed to handle - linking the BAM architecture directly to the connectivity innovation wave as well.
What the architecture could unlock is significant: any task that requires adapting physical actions to spoken or visual instructions becomes automatable at scale. Warehouse picking, assembly line quality checks, hospital logistics, elder-care assistance - all of these require precisely the combination of language understanding and dexterous physical response that BAM claims to provide. The granted patent expires in 2045: a 19-year window during which any competitor who copies this architecture must negotiate a licence or design around it.
Figure AI versus the competition: a patent race in slow motion
Figure AI holds 22 patents in the humanoid robotics space as of 2026 (GreyB data). Compare this to Ubtech Robotics at 521, Honda at 372, Toyota at 214, and Sony at 206. On raw patent count, Figure AI barely registers. But count is the wrong metric: the companies with the most filings have largely built their portfolios around mechanical design, hydraulics, and motion planning for robots designed before large AI models existed. Figure AI is filing much later, much more strategically, and on a software control layer that everything else depends on.
Boston Dynamics (now part of Hyundai Motor Group, with Atlas deployed at its Metaplant in Georgia) is the most physically capable competitor - but its IP concentration is in actuator design and dynamic motion, not AI language control. Tesla's Optimus program is the most visible threat on the AI control side, with Elon Musk projecting "thousands" of units in 2026 and millions by 2029. Unitree (China) shocked the industry with a $16,000 price point on its G1 model. What none of these competitors has, yet, is a granted US patent on a dual-model AI brain architecture. That gap - if Figure AI can maintain it - is worth far more than the 22-patent portfolio count suggests.
The global patent race: who is writing the robot's rulebook
The humanoid robotics patent landscape contains 19,501 patents filed between 2006 and 2026, across 15,022 unique families (GreyB, 2026). China accounts for roughly 9,975 of these - nearly four times the US total of 2,574. South Korea adds another 1,147. The volume tells one story: Asian manufacturers are filing intensively at the hardware and mechanical layer. But volume leadership in earlier generations rarely translates to platform control in the next one.
The year 2025 saw 2,592 new humanoid robotics patent families filed globally - more than double the 1,087 filed in 2023 (GreyB, 2026). Nearly 48% of all patent activity in this field has occurred in the last five years. The race is accelerating, and the most contested ground is now the AI control stack: the software layer that tells the hardware what to do. Countries and companies that win that layer collect royalties from everyone else. Countries that lead only on mechanical manufacturing are replaceable by the next production cycle. The BAM patent is a bid for the former position.
What does it mean for us?
Humanoid robots are no longer a laboratory curiosity. Figure 03 is already running on the floor of BMW's Spartanburg plant. TrendForce projects more than 50,000 humanoid robot units shipped globally in 2026 - a 700% surge over 2025. The global market is on a trajectory from roughly $3 billion today toward $15 billion by 2030, at a compound annual growth rate near 39% (MarketsandMarkets, 2025 data).
The BAM patent signals something specific: the critical competition is happening at the software control layer, not the mechanical one. The company that owns the architecture through which a robot understands language and acts on it owns a structural advantage that persists even as cheaper hardware proliferates. For technology companies, manufacturers, and policymakers outside the two main patent jurisdictions, the lesson is clear: engage with this IP now, not when it is already embedded in every production line. And for anyone who needs to file, enforce, or understand patents like US12578733B2 in non-English markets - Vietnam, Japan, Korea, the EU - precise patent translation is not optional. A mistranslation of a claim boundary can void years of protection or expose an infringement that was never intended.
| Field | Detail |
|---|---|
| Granted Patent No. | US12578733B2 |
| Application No. | US19/319,712 (pub. US20260064125A1) |
| Title | Bipedal action model for humanoid robot |
| Assignee | Figure AI Inc (California, USA) |
| Filing Date | September 4, 2025 |
| Publication Date | March 5, 2026 |
| Grant Date | March 17, 2026 |
| Jurisdiction | United States (USPTO) |
| Expiration | September 4, 2045 |
| Key Innovation | Dual-model AI system: alpha (language, 1-20 Hz) + beta (motor control, 100-10,000 Hz) |
FAQ
What does Figure AI's patent US12578733B2 actually cover?
The patent covers a "Bipedal Action Model" (BAM) - a dual-AI system where a large language-capable "alpha" model handles task planning at 1-20 Hz, while a fast "beta" model translates that intent into continuous motor commands for 62 joints at up to 10,000 Hz. Together they let the humanoid robot understand spoken instructions and move fluidly in real time.
Is this patent granted or still a pending application?
The patent was granted by the USPTO on March 17, 2026 (granted patent number US12578733B2). It was published 12 days earlier as application US20260064125A1 on March 5, 2026. The patent is currently active and expires in 2045.
How does a humanoid robot patent like this need technical translation?
Patents like US12578733B2 require accurate patent translation to be filed and enforced in non-English jurisdictions. A single mistranslated claim term - such as "continuous control commands" vs. "discrete control commands" - can narrow or void the scope of protection. Technical and IP translation by a specialist is essential for robotics patents entering Asian or European markets.
Who are Figure AI's main competitors in the humanoid robot space?
The primary competitors are Boston Dynamics (owned by Hyundai), Tesla Optimus, China's Unitree, and Agility Robotics. Each is racing to build both hardware and AI control software IP. Figure AI's strategic advantage is being one of the first to hold a granted US patent on a foundational AI control architecture, rather than only on mechanical design.
Why does the two-model alpha-beta design matter for the robotics industry?
Because it solves a fundamental incompatibility: language models think slowly, motor controllers must react fast. By separating these into two models on two GPUs, Figure AI's architecture lets the robot understand language and move smoothly at the same time - something older systems could not do. Any competitor that uses a similar design must now licence this patent or engineer a workaround.
Sources
Google Patents - US20260064125A1 / US12578733B2 (Figure AI, 2026)
Figure AI Series C Announcement - figure.ai (2025)
GreyB - Humanoid Robotics Patent Landscape 2026
MarketsandMarkets - Humanoid Robot Market Report 2025-2030
About the author
Dao Huy (Lucas) is a professional translator with over seven years of experience specialising in technical translation, patent translation, and IP translation across English, Chinese, and French into Vietnamese. He works regularly with robotics documentation, AI system specifications, and technology patents where precise terminology - and the ability to render a claim boundary faithfully across languages - determines the scope of legal protection. Patents like US12578733B2 are precisely the documents that need expert handling when entering Vietnamese or other Asian markets.
If your company holds or is acquiring robotics, AI, or technology patents and needs them translated and localised for Vietnam or the broader ASEAN region, reach out for a consultation at daohuy.com. Services include patent translation, engineering document translation, software localisation Vietnamese, and technology localisation for the Vietnamese market.
Written by Dao Huy (Lucas), Vietnamese translator & localization specialist (EN · ZH · FR → Vietnamese). See translation services →