Robot simulation software: simulators, synthetic data and open libraries
Robot learning in 2026 goes in two stages: first the robot trains in robot simulation software on synthetic data, then the model is fine-tuned on recordings from a real robot. For this you need a physics engine, a robot operating environment and a training library. Below are open and free tools, ordered by license openness and access from Russia, with data as of 16 October 2026.
In short
- Fast experiments with learning movement: MuJoCo
- Simulating a mobile robot together with ROS 2: Gazebo
- Controlling a real robot and sensors: ROS 2
- A training project with an open arm and a model: LeRobot and SO-101
- A photorealistic scene and humanoid training: Isaac Sim and Isaac Lab
- Synthetic video for training: Cosmos
The list
The order follows the criteria of this page; it is not a rating.
- #1
MuJoCo
Google DeepMind, open-source project
- What it is good for
- A physics engine for learning movement: walking, grasping, balance, research and study.
- 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
- #2
Gazebo
Open Robotics, open-source project
- What it is good for
- A simulator for mobile robots and manipulators that works together with ROS 2.
- Why it is on the list
- Gazebo Jetty came out in September 2025 as a long-term support version, supported until May 2031. Gazebo Docs, September 2025
- Price
- free (16 October 2026) plans
- Russian language
- no per the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- Apache 2.0
- Official site
- gazebosim.org
- #3
ROS 2
Open Robotics, open-source project
- What it is good for
- The robot’s operating environment: it links sensors, motors, navigation and models into one system.
- Why it is on the list
- On 22 May 2026, ROS 2 Lyrical Luth came out, the twelfth ROS 2 release, with long-term support until May 2031. ROS 2 Documentation, 22 May 2026
- Price
- free (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- ros.org
- #4
LeRobot and SO-101
Hugging Face, open-source project
- What it is good for
- A library for training on robot recordings and the open SO-101 arm for a training bench.
- Why it is on the list
- As of 16 October 2026, the LeRobot repository on GitHub had more than 28,000 stars; the latest version, 0.6.1, came out on 3 August 2026. GitHub, huggingface/lerobot, 16 October 2026
- Status
- The SO-101 kit is sold through partners from the official list: for example, as of 16 October 2026 RobotEd (Switzerland) shows the full kit in stock, shipping within two business days. source
- Price
- the library is free; parts for a pair of SO-101 arms (leader and follower) per the official repository’s bill of materials cost $229.88, one follower arm $121.94, excluding 3D-printed parts (16 October 2026) plans
- Russian language
- no per the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- github.com/huggingface/lerobot
- #5
Isaac Sim and Isaac Lab
NVIDIA, USA
- What it is good for
- Photorealistic simulation and reinforcement learning, including for humanoids.
- 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
- #6
Cosmos
NVIDIA, USA
- What it is good for
- World models: they generate video of scenes to add to the data from the robot’s cameras.
- 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
In short: a tool for each task
All six tools are free. The first four are open under Apache 2.0 and download with no country limits, so they rank higher. Isaac Sim is also free, but by its catalog it does not officially support Russia, and distributing Omniverse Kit in your own product needs an Omniverse Enterprise subscription.
The usual stack for a training or research project: MuJoCo or Gazebo for simulation, ROS 2 for control and LeRobot for training a model on recordings. Isaac Sim and Cosmos are needed when you want a photorealistic picture and a lot of 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..
Simulators
MuJoCo from Google DeepMind models body contacts and joints, so it is used to teach walking, balancing and picking up objects. Version 3.16.0 came out on 14 October 2026, and the GitHub repository had more than 15,700 stars by 16 October 2026. NVIDIA, Google DeepMind and Disney Research are developing the open physics engine Newton, and MuJoCo-Warp should 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.
Gazebo from Open Robotics suits mobile robots: warehouses, outdoor sites, lidars and cameras. The Gazebo Jetty version came out in September 2025 with long-term support until May 2031.
Isaac Sim and Isaac Lab from NVIDIA give a photorealistic scene and training of thousands of robot copies in parallel on a graphics card. On 16 September 2026, an early version of Isaac Lab 3.0 came out, where some training scenarios work without installing Isaac Sim; the general release is planned for the end of October 2026. Training in a simulator often goes by trial and error, and the glossary article on reinforcement learningreinforcement learning. Reinforcement learning (RL) is a way of training in which a program tries actions and learns from a reward for good steps. describes this method.
Synthetic data
Recordings from a real robot are expensive: you need a robot, an operator and a site. A simulator gives synthetic data: thousands of repeats of a task with different light, objects and noise.
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. A mix of synthetic and real data gave +40% to the result.
On 6 January 2025, NVIDIA released Cosmos, open 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. for training robots and self-driving vehicles on synthetic data. Among the first users were 1X, Agility, Figure AI, Uber and XPENG. On 1 June 2026, the Cosmos 3 family came out under the OpenMDW-1.1 license; the largest model has 64 billion parameters.
The main difficulty is called the gap between simulation and reality: in a simulator friction, light and the weight of objects are perfect, and on a real site they are not. So scene parameters are changed at random, so the model does not get used to one picture, and synthetic data is mixed with real recordings.
Real recordings are still needed. Open X-Embodiment from Google DeepMind and partner labs collected data from 22 types of robots, more than 1 million episodes and 500 skills: the RT-1-X model trained on it is on average 50% more successful than the original methods. What synthetic data is and where its limits are, the glossary explains.
The ROS 2 operating environment
ROS 2 is not an operating system in the usual sense but a set of programs on top of Linux, through which sensors, motors, navigation and models exchange messages. The main packages are open under Apache 2.0.
On 22 May 2026, ROS 2 Lyrical Luth came out, the twelfth ROS 2 release, with long-term support until May 2031. For a training project, take a long-term support release: drivers and examples come out for it.
Makers of industrial robots are also moving to ROS 2: in March 2026, FANUC released an official ROS 2 driver for its robots. This means that skills gained on a training bench will also be useful in a factory.
Simulators connect to ROS 2 directly: the same program controls both a virtual robot in Gazebo or Isaac Sim and a real one. So a logic error is caught in simulation, not on expensive equipment.
LeRobot and an open arm for $130
LeRobot from Hugging Face is an open library for training robots on recordings: record demonstrations, train a model, run it on the arm. As of 16 October 2026, the repository had more than 28,000 stars, and version 0.6.1 came out on 3 August 2026.
In April 2025, per Gigazine, the open SO-101 arm cost about $100–130 in parts and $240 as a ready kit. By the specification of the official repository as of 16 October 2026, the parts for a pair of arms (leader and follower) cost $229.88, and one follower arm $121.94, without the 3D-printed parts.
A person moves the leader arm, the follower repeats the movements, and LeRobot records image and movement pairs. Models such as SmolVLA or π0 are fine-tuned on these recordings; they are reviewed on the page on models for robots.
How to start a training project
Step one: install a long-term support release of ROS 2 and go through the basic examples with simulation in Gazebo. This shows how a robot works as a system.
Step two: train a simple motion policy in MuJoCo, for example keeping balance or grasping a cube. Here you see how trial-and-error learning works and how many attempts the model needs.
Step three: build an SO-101 arm, record a few dozen demonstrations of one task in LeRobot and 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. an open model on them. Compare the share of successful attempts before and after fine-tuning.
At each step, keep a log: the model version, the number of demonstrations, the scene conditions and the share of successful attempts. It shows what gave the result: new recordings, synthetic data or another model, and the project can be repeated a month later.
Courses and programs that teach this from zero are collected on the page on robotics and AI courses.
How we built this top
- Open license and free use
- Access from Russia per the official page
- Role in the training chain: simulation, control, learning from recordings, synthetic data
- Project activity: date of the last release and the support term
The order follows the criteria of this page; it is not a rating.
Checked:
How we build ratings