Figure AI Patent US20260064125A1: The Two-Brain Architecture Behind Useful Humanoid Robots
💡 Figure AI (US20260064125A1) just published a patent on a two-brain architecture for humanoid robots: one neural network reasons like a planner at 1-20 Hz, and a second controls every joint like a reflex at 100-10,000 Hz - both running on the robot's own GPUs, no cloud connection required. Filed September 4, 2025; published March 5, 2026. This is the clearest blueprint yet for on-device physical AI from the world's most valuable humanoid robotics startup, and it is the document that will shape robot IP battles for the next decade.
The bipedal action model (BAM) is emerging as one of the most consequential AI architectures of the decade, and US20260064125A1 - a patent application filed by Figure AI Inc. on September 4, 2025 and published March 5, 2026 - is the clearest public claim on record for how to build one. At stake is not just who wins the humanoid robot race; it is who owns the legal terrain on which that race is run.
What the Patent Actually Claims to Do
The invention describes a hierarchical two-model architecture built to fit inside a humanoid robot's torso. The first layer, the alpha model, receives natural language commands - "carry the box to shelf three" - and processes the visual scene through onboard cameras. It runs at a deliberate 1 to 20 Hz: the frequency of thoughtful planning. The second layer, the beta model, takes the alpha model's output and translates it into precise torques and velocities for every joint in the robot's body, operating at 100 to 10,000 Hz - the frequency of physical reflexes.
The two models are connected by a mechanism the patent calls action chunking: the alpha model packages its intent into short predicted movement sequences spanning 50 to 150 milliseconds, and the beta model executes those sequences as a smooth, continuous motor stream. The alpha model contains more parameters for reasoning; the beta model is leaner but faster for execution. Both run on separate GPUs located on the robot itself, not on a remote server. The training pipeline combines internet-scale video and text, robotic simulation, and human teleoperation of the physical robot - all within a single unified framework.
| Field | Detail |
|---|---|
| Patent number | US20260064125A1 |
| Title | Bipedal action model for humanoid robot |
| Assignee | Figure AI Inc. |
| Inventors | Corey Lynch, Yevgen Chebotar, Toki Migimatsu, Michael Ahn |
| Filed | September 4, 2025 |
| Published | March 5, 2026 |
| Jurisdiction | United States (USPTO) |
| Alpha model rate | 1-20 Hz (reasoning and planning) |
| Beta model rate | 100-10,000 Hz (joint-level control) |
| Action chunk size | 50-150 milliseconds of predicted motion |
These numbers are not incidental. They are the technical heart of the patent claim, and they bridge a gap that has quietly blocked the whole field of robotics for years.
The Problem This Architecture Solves
A large language model asked to plan the next step of a task thinks at roughly 1-10 queries per second. A robotic arm executing smooth motion needs positional updates 500 to 1,000 times per second. These two timescales are almost three orders of magnitude apart. Every robot that tried to close this gap through a remote-server connection had to accept the latency of the network in between.
Network latency does not sound fatal until a robot misses a grip because the command arrived 200 milliseconds too late, or until it loses connectivity in a warehouse basement and falls over. The BAM architecture dissolves this problem by co-locating both the planner and the controller on a single physical platform, designing them to communicate via an efficient, asynchronous chunk interface. The planner does not wait for each motor command to complete; the controller does not pause while the planner finishes each sentence.
The immediate consequence is a robot that reacts to changing scenes in real time while still following multi-step instructions - something no off-the-shelf system reliably achieves today, as the 40% success rate on "seal a ziplock bag" in the chart above makes plain. The longer-term consequence is that deploying robots in new environments becomes cheaper and faster, because the intelligence travels with the hardware. That changes the economics of robotics deployment at scale.
What This Invention Connects to and What It Could Unlock
The BAM patent sits at the intersection of three large technology fields. The first is AI inference hardware: the entire architecture depends on mobile GPUs small and powerful enough to fit inside a torso while running billions of parameters in real time. NVIDIA's edge AI chips and the wider race for high-efficiency inference hardware are direct enablers of what Figure AI is patenting. The second field is robot simulation and synthetic data: the training pipeline described in the patent requires generating millions of simulated robot experiences, which relies on physics simulators that companies like NVIDIA (Isaac Sim) are building in parallel. The third field is foundation AI models: the alpha model's language and visual understanding is built on the same transformer architectures as the largest language and vision models, meaning advances in those foundations directly raise the ceiling of what the alpha component can do.
Looking outward from this node, the BAM patent - if the claims are held broadly - could act as a toll road for competitors. Any humanoid robot with a two-tier onboard architecture, a reasoning model feeding an execution model through chunked action sequences, would potentially need to negotiate a license with Figure AI. That includes teams inside Google DeepMind (whose Gemini Robotics 2, announced July 30, 2026, uses a similarly structured split between a vision-language-action model and an embodied reasoning model), as well as dozens of Chinese humanoid startups scaling manufacturing rapidly through 2026. The enforceability across jurisdictions is where skilled patent translation becomes essential, because a poorly translated claim in a Chinese or European filing can make the protection meaningless before it is ever tested in court.
Who Is Behind This Patent and What It Says About the Race
Figure AI Inc. was founded in 2022 by Brett Adcock. By late 2025, the company's valuation had reached 39 billion US dollars, underpinned by a 675 million dollar Series B in February 2024 from Microsoft, NVIDIA, Intel, Amazon, Jeff Bezos, and OpenAI. By June 2026, the company was operating 740 robots in warehouses and logistics facilities - more than its 600 human employees.
The four named inventors are not junior researchers. Corey Lynch, Figure AI's Director of AI, spent seven years at Robotics at Google building some of the field's most-cited work on language-conditioned imitation learning, then joined Figure AI to build the Helix VLA system. Yevgen Chebotar, Toki Migimatsu, and Michael Ahn contributed to robot learning at Google Brain and DeepMind before joining Figure AI. The pedigree of the team and the specificity of the technical claims signal that this is not a defensive filing aimed at a broad technology class - it is a precise claim on a working system the company intends to commercialize and defend.
Competitors cannot easily work around the dual-model, action-chunking, onboard-GPU architecture without landing on alternative IP. That forces the field to innovate around the patent - which, paradoxically, is exactly what strong patents are supposed to do.
So What Does It Mean for Us?
Five years ago, "a humanoid robot that follows a natural language instruction" was a research demo. Today it is a product category worth tens of billions of dollars, and the key legal documents are patent filings like US20260064125A1. The BAM patent does not guarantee that Figure AI will win the robot race - patents protect an invention, not a market position. But it establishes the right to exclude competitors from a specific technical approach the company believes is central to making bipedal robots useful in the real world.
What comes next is predictable from history: filing in additional jurisdictions (the EPO, Japan, South Korea, China), licensing negotiations, and eventually litigation. At every step, the quality of the translation determines the quality of the protection. An "action chunk" mis-described in a Japanese filing, or a "torso-embedded GPU" rendered ambiguously in a Chinese claim, can shrink the scope of a patent worth hundreds of millions of dollars to something the competition can easily step over. The technical and legal precision required here is exactly where specialist patent translation earns its value.
FAQ
What is the Bipedal Action Model (BAM) in Figure AI's patent?
The BAM is a two-tier AI architecture for humanoid robots. An alpha model (1-20 Hz) processes language commands and plans the next move; a beta model (100-10,000 Hz) converts those plans into precise joint torques. Both run on dedicated GPUs inside the robot's torso, with no cloud connection needed. They communicate via action chunks: predicted movement sequences of 50-150 ms.
How is this different from existing robot control systems?
Most prior systems split cognition and control across a cloud server and a local controller, accepting network latency between them. The BAM co-locates both on the physical robot, eliminating that latency. It also uses a unified training framework combining internet-scale data, simulation, and teleoperation - rather than separate pipelines for language understanding and motor control.
Why does the 100-10,000 Hz rate for the beta model matter?
Smooth physical motion requires control commands at high frequency. Human muscle systems operate in this range. At lower frequencies, a robot's movements appear jerky or lag behind real-world events. The patent's claim that the beta model operates at up to 10,000 Hz signals an architecture designed for real-world deployment rather than just laboratory demonstrations.
Who are the main competitors affected by this patent?
Any company building a humanoid robot with an onboard hierarchical VLA-plus-controller architecture could be affected: Google DeepMind (Gemini Robotics 2, July 2026), Tesla (Optimus program), and a range of Chinese humanoid startups. Enforceability depends heavily on how broadly the claims are interpreted in each jurisdiction - making quality patent translation in each target country critical.
Why does this patent matter for patent translation services?
As Figure AI and its rivals race to file in the US, Europe, Japan, Korea, and China, each jurisdiction requires a technically precise translation. A mistranslation of a key term like "action chunking" or "alpha model inference frequency" can narrow or invalidate a patent claim worth hundreds of millions of dollars. Specialist patent translation for robotics and AI filings is a growing, high-stakes service area.
Sources:
USPTO Patent US20260064125A1, Google Patents | Figure AI, Wikipedia | Gemini Robotics 2, Google DeepMind, July 2026 | Figure AI Patents, GreyB, 2026
About the Author
Dao Huy (Lucas) is a professional translator with over 7 years of experience in technical, patent, and IP translation, working from English, Chinese, and French into Vietnamese. He has localized software platforms, engineering documentation, and patent filings for clients in technology, robotics, and advanced manufacturing - sectors where precise technical language is not just preferred but legally required.
If your team needs patent translation, technical IP translation, or software and technology localization into Vietnamese for the Asia-Pacific market, Dao Huy offers specialized services for these high-precision needs. Request a quote at daohuy.com.
Written by Dao Huy (Lucas), Vietnamese translator & localization specialist (EN · ZH · FR → Vietnamese). See translation services →