Robot Foundation Models Shift the Humanoid Race to Software

The brain race has overtaken the body race

For most of the past decade, people judged humanoid robots by their bodies: how they walked, how many degrees of freedom their hands had, and how well they survived a shove. Coverage through 2026 suggests that standard is outdated. The most important progress in robotics is now happening in the models that control the machines, and the companies that read this shift correctly are the ones setting the agenda.

The speed of releases alone makes the case. In June 2026, developers released 13 new embodied AI foundation models and world models, about one every 48 hours. One analysis described this as robotics competition moving from mechanical capability to software intelligence. The releases came from China's research institutes and big tech companies as well as Western labs. The Beijing Academy of Artificial Intelligence released world-model projects, and Alibaba released a Qwen-Robot family that uses alignment techniques for specific tasks instead of relying on raw scale.

In the past three months the pace has picked up further. Google DeepMind launched Gemini Robotics 2 on July 30. It is the company's first model that controls a humanoid's legs, torso, arms and hands under a single policy, and Apptronik's Apollo 2 was the demonstration robot.27 Physical Intelligence opened its models to outside developers through an API beta on October 13. Customers can upload their own data, fine-tune models and run cloud-hosted inference on their own robots.24 On the same day, a little-known lab called Nature Will announced Natura-1. It is a 50-billion-parameter mixture-of-experts model for dexterous hand control, pre-trained on more than a million hours of first-person human video.28

What changed technically

The main idea comes from large language models. Instead of programming a robot for each job, labs pre-train a general model on large and varied data and then adapt it cheaply. The best-known version is the vision-language-action (VLA) model, which takes in camera images and plain-language instructions and outputs motor commands.17 DeepMind's first Gemini Robotics model, built on Gemini 2.0, simply added physical actions as another output type. The company said it more than doubled performance on a generalization benchmark compared with other leading VLA models.13

The 2026 releases go further in three ways.

Whole-body control. DeepMind says Gemini Robotics 2 can coordinate walking and manipulation together, so a humanoid can reposition itself, crouch, bend and reach as one continuous action. Earlier versions focused mostly on the upper body.18

Portability across robot bodies. DeepMind says its on-device model can be adapted to a robot it has never seen with fewer than 200 examples.18 Its product page claims Gemini Robotics 2 can be adapted to any two-armed robot in a few hours.11 Skild AI's "omni-bodied" approach takes the idea further: it trained across a simulated set of roughly 100,000 robot bodies so the model has to learn general rules of movement instead of memorizing one machine.17

Learning from humans instead of robots. This may matter most. Robot teleoperation data is scarce and expensive, so labs are turning to video of people. VioLA, a humanoid policy described in October research, was trained on a pool of 140.6 million frames, 93.2 percent of them from humans. It reportedly reached 100 percent success on locomotion tasks without task-specific fine-tuning, compared with 16.7 percent or less for baselines such as Nvidia's GR00T N1.7.25 Natura-1 uses the same approach for hands.28 Xiaomi's open-source Xiaomi-Robotics-1 was trained on more than 100,000 hours of real manipulation trajectories, and the company says performance keeps improving with more data and larger models, with no sign of leveling off.29

If that scaling claim keeps holding up, it is the strongest reason to call foundation models the real breakthrough. It means robot capability can be bought with data and compute, as language-model capability was, instead of being engineered one actuator at a time.

The money agrees

Investors are acting on this view. Startups had raised $8.6 billion for humanoid robotics by mid-2026, according to Dealroom figures cited by Tech Funding News. That is already 1.8 times the total for 2025.1 Tracxn counts the numbers differently and reports $6.61 billion in humanoid equity funding through September, a 222 percent increase over the same period last year.7 The trackers disagree on the totals but agree on the direction, and the gap mostly reflects how each one defines the sector.

The biggest checks are increasingly going to the model layer. Skild AI raised $1.4 billion in January for its robot-agnostic Skild Brain.28 Physical Intelligence has raised about $1.07 billion and is reportedly negotiating another $1 billion at a valuation above $11 billion.1 Figure, the best-funded pure-play humanoid maker, has signed a deal with Nscale for up to 100,000 Nvidia Vera Rubin GPUs. The initial commitment is $3.5 billion, with plans to go past $6 billion, all to train its Helix model. CEO Brett Adcock has said the company is now limited mainly by data and compute.28 Reports differ on how much Figure has raised in total: about $1.9 billion in one account, $2.34 billion in another, and $1.75 billion in a third.2817 Every account agrees that its biggest new spending is on compute, not factories.

The companies that build hardware describe themselves in model terms too. When UK-based Humanoid raised a $152 million Series A at a $1.35 billion valuation in July, it listed expanding its proprietary AI platform, KinetIQ, alongside its next-generation robot and the start of mass manufacturing.23 Reuters described the round as part of investor demand for "physical AI" companies.5

Where the thesis is weaker than it sounds

The claim that models matter more than hardware is mostly correct. It still needs two qualifications.

First, the models are not yet reliable. DeepMind's own results for Gemini Robotics 2 show whole-body picking success of about 68.4 percent from a table, 76.3 percent from a shelf and 45.7 percent from the floor. One logistics analyst said this makes it a research system, not a production replacement for warehouse automation.18 General frontier models have the same problem. In independent tests, a GPT-6 variant succeeded on 19 of 20 simple block-in-bowl attempts but only 2 of 20 precision insertions.28 A survey of the field through August found that two competing designs, VLA models and "world-action" models, both score in the high 90s on the LIBERO benchmark.21 Researchers are also warning about an "embodiment gap": the real test is whether a model can be deployed and adapted to physically different robots, which benchmark scores do not measure.26 High simulation scores and shaky real-world numbers suggest the benchmarks are saturating faster than real capability is improving.

Second, hardware still matters. Humanoid's round brought in Bosch as its contract manufacturer, responsible for hardware design, production and supply chain.6 Some observers argue that production volumes, such as Figure shipping more than 1,000 units and AgiBot reporting 15,000 cumulative units in the field, are a better measure of progress than demos.9 Deployed robots also produce the real-world data that models depend on. Tesla, for example, is using Optimus internally partly to gather training data.22

The commoditization problem

The most interesting disagreement is not between hardware and software. It is about which layer of software will capture the value. General Robotics CEO Ashish Kapoor describes three layers: frontier models at the bottom, robot foundation models in the middle, and task-specific post-training on top. He expects the first two layers to merge, which would squeeze the middle layer that has drawn billions in investment.28 Unitree founder Wang Xingxing adds to that argument. After testing GPT-6 on Unitree robots, he shortened his forecast for an embodied-AI tipping point from two to three years to one to two years.28

The evidence supports Kapoor more than the valuations suggest. A small lab can now release a 50-billion-parameter dexterity model.28 Xiaomi is open-sourcing its model.29 DeepMind is making its embodied-reasoning model, ER 2, available through the public Gemini API.19 When model capability spreads this quickly, the advantage usually moves to whoever controls proprietary deployment data and the work of turning one robot into a useful worker at one specific job. Physical Intelligence's API launch can be read as a response to this: it is trying to become a platform before the model itself becomes a commodity.2224

The bottom line

Foundation models are the real breakthrough in robotics. They are why a single brain can now run very different robots, why human video can stand in for expensive robot data, and why capital is moving toward compute. The breakthrough is still incomplete: floor-level picking succeeds less than half the time, and precision tasks still defeat general models. The companies best placed to benefit are probably not those with the largest models. They are the ones that pair good-enough models with robots in the field producing proprietary data. Hardware no longer leads the race, but it still decides which companies can collect that data.