Lerrel Pinto co-founded Fauna Robotics, a startup that built an approachable bipedal robot called Sprout. He left in 2025. Amazon acquired Fauna in March, along with its 50 employees and its $50,000, three-and-a-half-foot-tall dancing humanoid, to enter the consumer robotics market. Pinto then co-founded Assured Robot Intelligence with Xiaolong Wang, a former Nvidia researcher and associate professor at UC San Diego who won the MLSys 2024 Best Paper Award for work on AI model optimisation. On Friday, Meta acquired ARI and both founders joined Meta Superintelligence Labs. The acquisition closed the same day it was announced. Financial terms were not disclosed. The interesting question is not what Meta paid for a startup whose employees were concentrated in San Diego and New York. It is what Meta intends to do with the technology, and what that intention reveals about the company’s theory of how the humanoid market will develop.

The platform

Meta’s stated goal for robotics is to replicate what Google’s Android operating system and Qualcomm’s chips did for the smartphone industry: build the foundation that everyone else builds on. The company launched Meta Robotics Studio last year, hired former Cruise CEO Marc Whitten to lead the effort, and began recruiting roughly 100 engineers to develop in-house humanoid hardware alongside the AI models that power it. CTO Andrew Bosworth has described humanoid robots as Meta’s next bet of comparable scale to augmented reality, a category in which Meta has already spent tens of billions through its Reality Labs division. The ARI acquisition adds a specific capability to this effort: robot control models that enable humanoids to understand, predict, and adapt to human behaviour in unstructured environments.

The platform strategy is explicit. Meta intends to develop sensors, software, and AI models for robots and make them available to the rest of the industry, meaning the technology could be used by manufacturers that Meta does not own or control. This is the Android model applied to physical machines. In smartphones, Google gave away the operating system and captured value through search, advertising, and the Play Store ecosystem. In robotics, Meta would give away the intelligence layer and capture value through the data, the model ecosystem, and the integration with Meta’s existing platforms, where 3.3 billion people already interact daily. Meta has been acquiring AI talent aggressively, hiring five founding members of Thinking Machines Lab, including a researcher whose six-year package reportedly reached $1.5 billion. The ARI acquisition fits the same pattern: small team, frontier capability, immediate integration into the Superintelligence Labs research division.

The technology

ARI’s technical contribution centres on what the company calls “robotic intelligence designed to enable robots to understand, predict and adapt to human behaviors in complex and dynamic environments.” In practice, this means AI models for whole-body humanoid control, the ability to coordinate a robot’s limbs, balance, and movement in response to real-time sensory input from an unpredictable physical world. Wang’s award-winning work on activation-aware weight quantisation, the same technique that made Nebius’s $643 million acquisition of Eigen AI valuable this week, is relevant here: compressing AI models so they run efficiently on the limited compute available inside a robot, rather than requiring a connection to a remote data centre.

The company also developed e-Flesh, a tactile sensor that measures deformations in 3D-printable microstructures using magnets and magnetometers. Tactile sensing is one of the unsolved problems in humanoid robotics. A robot that can see its environment through cameras and lidar still cannot feel the difference between gripping an egg and gripping a tennis ball without tactile feedback. The gap between how robots learn in simulation and how they perform in the physical world remains the central obstacle to deployment at scale. ARI’s work on self-learning for robot control, combined with its sensor technology, addresses both sides of that gap: better models and better sensory input.

The market

The humanoid robotics market has moved from speculative to competitive in the span of 18 months. Tesla plans to begin large-scale production of its Optimus V3 humanoid between July and August, with annual capacity targets of one million units by late 2026 and a price point between $20,000 and $30,000. 1X Technologies has opened a factory in Hayward, California, to produce 10,000 NEO humanoid robots in its first year, with first-year production capacity selling out within five days of preorders opening. Apptronik has raised $520 million at a $5 billion valuation, partnering with Google DeepMind and its Gemini Robotics models. Amazon has made two robotics acquisitions in a single month. Unitree is targeting 20,000 humanoid shipments in 2026. Morgan Stanley projects the global humanoid robot market will reach $38 billion by 2035 and $5 trillion by 2050.

The competitive dynamics are clarifying into three tiers. The first tier is vertically integrated manufacturers, companies like Tesla and 1X that design, build, and sell the complete robot. The second tier is platform providers, companies that supply the intelligence layer, the operating system, or the key components that multiple manufacturers use. The third tier is the component suppliers, chipmakers and sensor companies that sell to both. Meta is positioning itself in the second tier, and it is not alone. Google, through DeepMind’s Gemini Robotics programme and its partnership with Apptronik, is pursuing a similar platform strategy. Europe is developing its own approach to the humanoid race, with companies and research institutions pursuing strategies that emphasise safety, industrial precision, and regulatory compliance over the speed-to-market approach favoured by American and Chinese competitors.

The bet

Meta’s history with hardware platforms is instructive. The company missed mobile. Facebook Home, its 2013 attempt to become the default interface on Android phones, was discontinued within a year. The company then spent more than $50 billion on Reality Labs attempting to own the next computing platform through virtual and augmented reality, a bet that has yet to produce returns at anything approaching the scale of its advertising business. The Ray-Ban Meta smart glasses are the closest the company has come to a successful hardware product outside of its core social media platforms, and even those are essentially an accessory for Meta’s AI assistant rather than a standalone computing device.

The robotics bet is different in one respect. Meta is not attempting to manufacture the hardware at scale itself. It is attempting to provide the intelligence, the models, the sensor technology, and the software stack, and let others build the machines. This is a lower-capital, higher-leverage strategy than the Reality Labs approach, and it plays to Meta’s genuine strengths in AI research, open-source model distribution, and platform economics. But it depends on the humanoid market developing the way the smartphone market developed: with hundreds of manufacturers needing a common software platform. If the market instead consolidates around a few vertically integrated players, each with proprietary AI, the Android model does not apply. Tesla is not looking for an operating system. Neither is 1X. The companies that might want Meta’s intelligence layer are the ones that do not yet exist, the humanoid equivalents of Samsung and Xiaomi and Oppo, manufacturers that can build bodies but need someone else to provide the brain. Meta is betting those companies are coming. The ARI acquisition is the latest investment in making sure that when they arrive, Meta’s technology is what they reach for first.