What is physical AI: how neural networks learn to act in the world
Physical AI is a set of neural networks that see the world through cameras and sensors, understand a command and act through a robot’s hands or a machine’s wheels. They learn from three sources: world models, synthetic data from simulators and recordings of real robot movements. Below are the main models and tools as of 16 October 2026, with licenses and source links.
In short
- World models for synthetic data: Cosmos from NVIDIA
- A simulator for training robots: Isaac Sim and Isaac Lab
- An open physics engine: MuJoCo from Google DeepMind
- An open model for humanoids: Isaac GR00T
- An open model for robot arms: π0
- A closed model with a planner through an API: Gemini Robotics
The list
The order follows the criteria of this page; it is not a rating.
- #1
Cosmos
NVIDIA, USA
- What it is good for
- World models: generate video of believable scenes to train robots and self-driving vehicles.
- Why it is on the list
- On 1 June 2026, at GTC Taipei, NVIDIA released Cosmos 3, an open family of world models; the largest model, Cosmos3-Super, has 64 billion parameters. NVIDIA, 1 June 2026
- Price
- free: open weights (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- OpenMDW-1.1 (Cosmos 3)
- Official site
- nvidia.com/en-us/ai/cosmos
- #2
Isaac Sim and Isaac Lab
NVIDIA, USA
- What it is good for
- Simulation and training: a virtual workshop or warehouse where a robot trains before it goes into the real world.
- Why it is on the list
- On 16 September 2026, an early version of Isaac Lab 3.0 came out, in which some training scenarios run without installing Isaac Sim; general release is planned for the end of October 2026. GitHub, isaac-sim/IsaacLab, 16 September 2026
- Price
- free; distributing Omniverse Kit in your own products requires an Omniverse Enterprise subscription (16 October 2026) plans
- Russian language
- no per the site
- Access from Russia
- not officially supported official page
- Official site
- developer.nvidia.com/isaac/sim
- #3
MuJoCo
Google DeepMind, open-source project
- What it is good for
- Physics engine: fast experiments with learning movements in research and teaching.
- Why it is on the list
- On 14 October 2026, MuJoCo 3.16.0 came out, and as of 16 October 2026 the project’s GitHub repository had more than 15,700 stars. GitHub, google-deepmind/mujoco, 14 October 2026
- Price
- free (16 October 2026) plans
- Russian language
- no per the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- mujoco.org
- #4
Isaac GR00T
NVIDIA, USA
- What it is good for
- A vision-language-action model for humanoids, open weights with the right to commercial use.
- Why it is on the list
- On 16 March 2026, NVIDIA presented Isaac GR00T N1.7, an open model for humanoid robots licensed for commercial use; Humanoid, LG Electronics, NEURA and Noble Machines announced that they use it. NVIDIA Newsroom, 16 March 2026
- Price
- free: open weights and code (16 October 2026) plans
- Official site
- developer.nvidia.com/isaac/gr00t
- #5
π0 (openpi)
Physical Intelligence, USA
- What it is good for
- An open model for robot arms: fine-tuning on a new task from robot recordings.
- Why it is on the list
- On 4 February 2025, Physical Intelligence opened the code and weights of the π0 and π0-FAST models; according to the company, 1 to 20 hours of data were enough to fine-tune them for a new task. Physical Intelligence, 4 February 2025
- Price
- free: open code and weights (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- github.com/Physical-Intelligence/openpi
- #6
Gemini Robotics
Google DeepMind, USA
- What it is good for
- A closed model: the ER 2 planner through an API, action models for early-access partners.
- Why it is on the list
- On 30 July 2026, Google DeepMind released Gemini Robotics 2: according to the company, the model controls a whole humanoid robot and adapts to a new robot from fewer than 200 examples. Google DeepMind, 30 July 2026
- Price
- Gemini Robotics ER 2 in the Gemini API: free plan; paid $1 per 1 million input tokens and $5 per 1 million output tokens until 31 December 2026. The Gemini Robotics 2 and On-Device 2 action models are available only to early access partners (16 October 2026) plans
- Access from Russia
- not officially supported official page
- Official site
- deepmind.google/models/gemini-robotics
What is physical AI?
An ordinary language modellarge language model. A large language model (LLM) is a neural network trained on a huge amount of text. It writes, summarizes and analyzes text on request. answers with text. Physical AI answers with an action: turn the gripper, take a step, brake. For this, a neural network has to understand geometry, weight, friction and the results of its movements.
The main difficulty is data. There is a lot of text to train language models, but recordings of a robot picking up a cup are few and expensive: each hour of data needs a robot, an operator and a site. So physical AI is built from three parts: world models, simulation with synthetic datasynthetic data. Synthetic data is examples created by a program or a neural network, not collected from real life: texts, tables and images used to train and test models., and vision-language-action models that turn a command into movement.
One of the first open platforms for this was Cosmos: NVIDIA released it on 6 January 2025 to train robots and self-driving vehicles. Robots that already work for customers are covered in detail in the robotics review.
World models
A world model predicts how a scene will change after an action: where a box will roll, what a car’s camera will see in a second. With such a model, you can test thousands of options without the risk of breaking a robot.
Cosmos from NVIDIA came out on 6 January 2025 under the NVIDIA open license on Hugging Face. Among the first users, the company named 1X, Agility, Figure AI, Uber and XPENG. On 1 June 2026, at GTC Taipei, the Cosmos 3 family came out; its largest model, Cosmos3-Super, has 64 billion parameters.
How a world model works and how it differs from a video generator is explained in more detail in the glossary.
Synthetic data and simulation
Synthetic data are examples created by a computer rather than recorded from a real robot. They are generated in a simulator: a virtual robot repeats a task thousands of times with different light, objects and noise.
The effect is confirmed by numbers. On 18 March 2025, NVIDIA reported that in 11 hours it generated 780,000 synthetic trajectories for the GR00T N1 model, equal to 6,500 hours of human demonstrations, and a mix of synthetic and real data improved the result by 40%.
Simulators are getting faster. NVIDIA, Google DeepMind and Disney Research are developing the open physics engine Newton, and MuJoCo-Warp is expected to speed up robot machine learningmachine learning. Machine learning (ML) is a field of AI in which a program finds rules in data by itself instead of getting them ready-made from a developer. tasks more than 70 times. A comparison of simulators and libraries is on the page about robot training.
Vision-language-action models
A vision-language-action model takes camera frames and a spoken or written command and returns joint movements. Inside there are often two parts: a slow one plans the steps, a fast one controls the motors. That is how GR00T N1 from NVIDIA works.
On 4 February 2025, Physical Intelligence released the code and weights of the π0 model: according to the company, 1–20 hours of robot data are enough to fine-tune it for a new task. The authors themselves call the release an experiment.
On 25 September 2025, Google DeepMind presented Gemini Robotics 1.5, which thinks before it moves and explains its reasoning, and the Gemini Robotics-ER 1.5 planner. Skills learned on one ALOHA 2 robot carry over to the Apollo humanoid and the Franka arm. We compare open and closed models of this class on the page about models for robots.
Who builds physical AI
NVIDIA covers the whole chain: the Cosmos world modelsworld model. A world model is an internal model of the environment inside a neural network. It predicts how the world will change after an action and lets an agent plan before it acts., the Isaac Sim simulator with the Isaac Lab library, and the GR00T model. On 16 March 2026, the company released GR00T N1.7 with the right to commercial use, and Humanoid, LG Electronics, NEURA and Noble Machines announced they would use it. On 16 September 2026, an early version of Isaac Lab 3.0 came out, where some training scenarios work without installing Isaac Sim.
Google DeepMind develops Gemini Robotics and the open MuJoCo engine: MuJoCo 3.16.0 came out on 14 October 2026. Back in 2023, DeepMind and partner labs built Open X-Embodiment, an open datasetdataset. A dataset is a set of data collected and formatted to train or test a model: a table, a folder of pictures, or a body of texts in which every example is described in the same way. from 22 types of robots with more than 1 million episodes and 500 skills.
On 30 July 2026, Google DeepMind released Gemini Robotics 2: according to the company, the model controls a whole humanoid and adapts to a new robot from fewer than 200 examples. The action models are so far available only to early-access partners, while the ER 2 planner is open to developers through the Gemini API.
Physical Intelligence bets on the open π0 models. Robot makers, from Figure AI to 1X, build their own models or fine-tunemodel fine-tuning. Model fine-tuning is the adjustment of an already trained model on your own examples, so that it performs a specific task more accurately or writes in the required style. open ones.
Where business can use it
The nearest value of physical AI is not in humanoids but in tasks where a camera or a robot arm already stands: grasping parts without precise placement, quality inspection, warehouse route planning, retooling a line for a new product.
Simulation is useful even without a robot. A digital copy of a site in Isaac Sim or MuJoCo lets you test the layout and movements before buying equipment. Note the limits: according to the catalog, Isaac Sim does not officially support Russia, while MuJoCo, GR00T, π0 and Cosmos download as open projects with no country limits.
A physical AI pilotAI pilot project. An AI pilot project is a short launch of an AI solution on one process, with a metric set in advance, to check the effect before a large investment. usually goes in three steps. First, choose one operation and record how it is done now: how much time, how many defects. Then test the task in a simulator on open tools. Only after that, install the equipment on site and compare the numbers before and after.
Open weights can run on your own server, so video from the site does not go abroad. How such a setup works is covered in the section on sovereign AIsovereign AI. By definition, sovereign AI is neural networks and infrastructure that a country or a company controls itself: development, servers and data stay with the owner, not with a foreign provider..
How we built this top
- Role in the physical AI chain: world model, simulator, action model
- Openness: whether you can download the weights or code and use them in a commercial project
- Confirmed results on real robots, per the developer, with a date
- Access from Russia per the official page
The order follows the criteria of this page; it is not a rating.
Checked:
How we build ratings