AI Robotics Companies: Who Builds the Robot Brains

What an AI Robotics Company Actually Sells

An AI robotics company sells one of two things: the software that decides what a robot does, or the machine that does it. The label covers both, which is why the category looks crowded and confusing.

The first group builds robot foundation models. These are large models trained on robot and human demonstration data, and they output motor actions instead of text. The company ships software, not steel.

The second group builds hardware and trains its own model to run on it. Figure, 1X and Tesla all sit here. They control the whole stack, so their model only has to work on one body.

The distinction matters for buyers. A brain company can reach you through a robot arm you already own. A body company cannot sell you anything until its robot ships.

For the underlying concept, see what embodied AI is and what AI robots are. This page covers who is building them.

The Robot Brain Companies (Foundation Models)

Robot foundation model companies build hardware-agnostic software meant to control many different robot bodies. None of them sell a robot to the public today, and most sell to hardware makers and industrial operators instead.

Physical Intelligence is the San Francisco lab behind the Pi0 and Pi0.5 vision-language-action models. Founded in 2024, it has raised more than $1.1 billion across three rounds and was valued near $5.6 billion in its November 2025 round, with a further round reported in 2026 targeting about $11 billion. Its models are designed to run on many robot bodies rather than one.

Skild AI was founded by former Carnegie Mellon professors Deepak Pathak and Abhinav Gupta and is based in Pittsburgh. Its Skild Brain is pitched as an omni-bodied model that controls different robots without retraining for each one. In January 2026 it raised a $1.4 billion Series C led by SoftBank, valuing the company above $14 billion.

Field AI takes a narrower target. Its Field Foundation Models are built for robots working in unstructured, GPS-denied places without a pre-built map. The company has raised roughly $405 million at about a $2 billion valuation, and its autonomy stack runs on Boston Dynamics Spot fleets.

Dyna Robotics is the most commercially blunt of the group. Its DYNA-1 model runs stationary arms on repetitive tasks, and the company has shown it folding napkins for 24 hours without human intervention. It has raised $144 million, including a $120 million Series A in September 2025.

Two more are worth knowing. Generalist AI trains its GEN-1 model largely from wearable-device demonstrations and runs it on existing collaborative arms. Covariant built RFM-1 for warehouse picking before Amazon hired its three founders in 2024 and licensed its models.

Genesis AI is the newest of these, based in Paris and San Carlos. It emerged from stealth in 2025 with a $105 million seed round and unveiled Eno in June 2026, a wheeled, non-humanoid robot with dexterous hands running its GENE model. It is the one company here that crosses the line, building both the model and the body.

The Body Builders Training Their Own Models

Body builders design the robot and its control model together, then keep both in-house. Their advantage is control, because the model never has to generalize to hardware the team did not build.

Figure, 1X, Tesla, Apptronik and Sanctuary AI all follow this pattern with humanoids. Agility Robotics does the same with Digit, its bipedal warehouse robot. Boston Dynamics splits the difference by running its own software on Atlas while partnering with Field AI on Spot.

Unitree Robotics and AgiBot pursue a third path, selling capable hardware cheaply and letting researchers supply much of the intelligence. That has made their machines the default platform in university labs.

None of these companies sell a general-purpose humanoid to consumers yet. Access runs through enterprise pilots and research agreements, which is covered in more detail in our humanoid robot companies guide.

The tell for a body builder is a named robot with published specs. The tell for a brain builder is a named model with published benchmarks and no robot of its own.

• Figure, 1X, Tesla, Apptronik, Sanctuary AI: humanoid hardware plus in-house models
• Agility Robotics: Digit, a bipedal robot aimed squarely at warehouse work
• Boston Dynamics: own software on Atlas, partnered autonomy on Spot
• Unitree Robotics, AgiBot: low-cost platforms that researchers program themselves
• LG: CLOi service robots, plus a bipedal humanoid that builds on NVIDIA Isaac GR00T and Jetson Thor rather than a wholly in-house model

Brain vs Body: Which Reaches a Buyer First

Brain companies reach paying customers sooner, because they attach to hardware that already works. Body companies carry the bigger long-term prize and the longer wait.

Dyna Robotics is the clearest example of the short path. The company has shown a pair of stationary arms folding napkins for more than 24 hours without a human resetting them, and that needs no new robot to exist.

A humanoid that walks into an unmodified building and does arbitrary work is worth far more. It also depends on hardware reliability, battery life and safety certification that are still open problems.

Choose by what you need automated. If a fixed arm can reach the work, a brain company can help this year. If the work requires moving through a human building, the timeline is measured in years.

How to Tell a Real AI Robotics Company From a Label

Almost every robotics company now calls itself an AI robotics company, so the label carries no information on its own. Four checks separate the companies with a real model from the ones with a marketing page.

Ask what the model is named and what it runs on. A company with a genuine foundation model names it, publishes what it controls, and says whether it works on hardware the company did not build.

Ask whether the demo was teleoperated. Remote human control looks identical to autonomy in a video, and the difference is the entire product. Dyna Robotics leans on the 24-hour unattended run for exactly this reason.

Ask who can buy it. A company with no purchase path for anyone outside a pilot program is a research lab with a funding round, which is a fair thing to be but not a supplier.

Across the 40 company profiles and 135 robot model pages The Bot Scout maintains, the pattern that separates the two groups is consistent. Companies with a real model publish the model name, the body it controls and a benchmark; companies without one publish a funding total and a render. Funding size tracks neither shipping product nor deployment count, which is why we treat a named model plus a named deployment as the qualifying evidence rather than the raise.

• Is the model named, and is it documented separately from the robot?
• Does it run on hardware the company did not build?
• Was the demo autonomous, or teleoperated by a remote human?
• Is there a purchase or licensing path outside an invite-only pilot?

Where the Funding Says the Field Is Going

The largest recent rounds went to companies selling software, not robots. Skild AI at over $14 billion and Physical Intelligence near $5.6 billion both build models that run on other people's hardware.

That weighting reflects a bet on reuse. A model that controls many robot types earns from every hardware maker that adopts it, while a robot company earns only from units it ships.

Backers overlap heavily across these rounds, with NVIDIA appearing as an investor in both Skild AI and Field AI while also supplying the compute and the Isaac and GR00T software many of them build on.

Treat valuations as expectations, not revenue. These are private marks set by one investor group, not a price any market has tested. How robotics firms reach public markets at all is a separate question, covered in our guide to robotics SPACs and IPOs.

Start with our company directory to see profiles, robot models and funding for each of these firms, then read what humanoid robots are if the hardware side is what you are evaluating.