Learning Objectives
- Understand the split between robot foundation models and the robots themselves, and why the models are no longer the bottleneck
- Explain the data problem that makes robotics fundamentally harder than text or image AI
- Read deployment claims critically — the difference between a demonstration, a pilot, and a robot doing paid work unattended
What Is Embodied AI?
Embodied AI is artificial intelligence that acts on the physical world rather than producing text or pixels. A chatbot's mistake is a bad paragraph you can ignore. An embodied model's mistake moves an object, or a person, or a two-hundred-pound machine — which is why almost every hard problem in this category is about reliability rather than capability.
The category divides in two, and keeping the halves separate is the single most useful thing you can do when reading news about it.
💡Key Concept
Two halves, often confused. A robot foundation model is software — weights that take in camera images and an instruction and output motor commands. A robot is the hardware those weights run on. They are made by different companies on different timelines, and a spectacular demo usually showcases one while quietly depending on the other. When a humanoid folds laundry in a video, ask which half is being demonstrated.
The Two Halves
| Half | What it is | Who makes it | Examples |
|---|---|---|---|
| Robot foundation models | Weights that turn camera input plus an instruction into motor commands | AI labs | Gemini Robotics 2, pi-0, Generalist GEN-1.5, Qwen-Robot, Robostral Navigate |
| The machines | Humanoid and mobile robot hardware | Robotics companies | Tesla Optimus, Figure 03, Boston Dynamics Atlas, Agility Digit, Unitree G1, Neura 4NE-1 |
| Simulation and world models | Environments that generate training data instead of collecting it | Mostly chip and platform vendors | NVIDIA Isaac and Omniverse, SANA-WM |
The interesting change of the last two years is in the first row. A robot foundation model used to be trained for one machine doing one job. Physical Intelligence's pi-0 controls seven different robot bodies across more than sixty manipulation tasks from a single set of weights, and Generalist's GEN-1.5 can pick up a new manipulation task from a demonstration of three to twelve seconds without retraining. Generalization across bodies, which is what the word foundation is doing in "robot foundation model", is real.
The Constraint Is Data, Not Intelligence
Here is the thing that makes robotics genuinely different from every other category in this directory. Language models were trained on a corpus that already existed — the public web. There is no equivalent corpus of robot behaviour. Nobody has published a trillion examples of a hand picking up a mug at slightly the wrong angle and recovering. That data has to be created, one demonstration at a time, by a person operating a real machine.
That single fact explains most of what the field is doing:
- Simulation is the bet. NVIDIA's Isaac and Omniverse stack exists so training runs can happen in a physics simulator millions of times faster and cheaper than in reality. The open question is the reality gap — how much a policy learned in simulation degrades on a real machine.
- World models are the newer bet. A world model learns to predict what happens next in video, so it can generate plausible experience rather than replay recorded experience. NVIDIA's open-source SANA-WM is an early public example.
- Teleoperation is the quiet workhorse. Many "autonomous" robots ship with a remote human fallback, and those human corrections are themselves valuable training data. Weave Robotics is unusually straightforward about this — Isaac 1 is autonomous by default with remote teleoperation behind it.
⚠️Warning
Read deployment claims carefully — the words do a lot of work. "Deployed" can mean a robot doing paid work unattended, or a pilot with engineers present, or a unit sitting in a customer's building. A demonstration video is edited, often teleoperated, and says nothing about reliability across a shift.
Three questions separate real deployments from good marketing: How many units, at how many customer sites, doing what task? Is a human intervening, and how often? Is the robot paid for by the hour, or provided free as a pilot? Agility's Digit is a useful benchmark here because it is sold on a service contract — someone is paying for output rather than for a machine.
Why This Category Matters Now
Two things changed at once. Capable multimodal models arrived that can look at a scene and reason about it, and humanoid hardware got cheap enough to buy in quantity — Unitree's lineup in particular has made it the highest-volume humanoid maker by pushing prices down rather than capability up. The result is that the field's constraint moved from "can a model understand the task" to "can we get enough physical data, and will it work on the tenth-thousandth try."
That reframing is worth carrying into any robotics story you read. Progress on benchmarks is now much easier than progress on reliability, and the companies talking about reliability are usually further along than the ones showing the most impressive video.
The Tools Worth Knowing
| Tool | Best For |
|---|---|
| Gemini Robotics 2 | Google DeepMind's embodied family — an embodied reasoning model that plans tasks plus a vision-language-action model that executes them |
| pi-0 | Physical Intelligence's general-purpose robot foundation model — one set of weights driving seven robot embodiments across 60-plus tasks |
| NVIDIA Isaac & Omniverse | The dominant simulation and training stack, used by Figure, Agility and most of the humanoid field to generate data cheaply |
| Tesla Optimus | The most vertically integrated humanoid programme — Tesla builds the robot, the chips and the model, and runs them in its own factories |
| Figure 03 | Third-generation humanoid from Figure AI, with the BMW deployment that is among the most-cited commercial humanoid installations |
| Boston Dynamics Atlas | All-electric humanoid with 56 degrees of freedom, moving from research showcase into warehouse and logistics work |
| Agility Digit | A bipedal warehouse robot sold on a service contract — the clearest example of a humanoid being paid for by the hour |
| Unitree G1 | Compact research humanoid from the highest-volume maker, and the machine most likely to be in a university lab |
| SANA-WM | Open-source video world model from NVIDIA Labs, Apache 2.0 — a public example of generating training experience rather than recording it |
The catalog also carries the newer model labs — Generalist, Genesis AI, Qwen-Robot, Robostral Navigate and Assured Robot Intelligence — plus applied systems like Gritt Robotics in construction and Weave Robotics in the home.
How to Think About Robotics & Embodied AI
If you are evaluating the field rather than buying a robot, track the data story rather than the demo reel: who has real-world hours, who is generating experience in simulation, and who is honest about teleoperation. If you are a developer, the accessible entry point is not a humanoid — it is the simulation stack and the open models, which run on hardware you already have. And if you are reading a funding announcement, remember that valuations here are being set on a bet about the next decade, not on current revenue.
Key Takeaways
- Embodied AI splits into robot foundation models and the machines they run on — different companies, different timelines, and a demo usually showcases one while depending on the other
- The models generalize now: pi-0 drives seven robot embodiments from one set of weights, and Generalist's GEN-1.5 learns a new manipulation task from a demonstration of a few seconds
- The binding constraint is data, not intelligence — there is no public corpus of robot behaviour the way there is for text, so every demonstration has to be created
- That is why simulation (NVIDIA Isaac and Omniverse) and world models (SANA-WM) sit at the centre of the field, and why teleoperation quietly supplies both a safety net and training data
- Read deployment language critically — ask how many units, at how many sites, doing what task, with how much human intervention, and whether anyone is paying for the output
- Progress on benchmarks is now far easier than progress on reliability, so the companies discussing reliability are often further along than the ones with the best video
