From your first experiment to an existing industrial cell. Compare robots across manufacturers, check your work area in 3D, and plan training, deployment and cost with LeRobot.
Start with your situation, then refine the setup. You can explore every step without an account.
Choose the goal and task below. The remaining steps update as you go.
IL: a good starting point when people can demonstrate the behavior. Record examples, train ACT or fine-tune a VLA, evaluate and collect corrections.
RL: consider it when you can score success and repeat trials with resets. This advisor models HIL-SERL with supervised robot practice, a reward classifier and human interventions.
For example: demonstrate varied pick-and-place tasks with IL; consider RL for a repeatable insertion task with a reliable reward. Precision alone does not decide the method. Demonstrations can also bootstrap RL.
Choose a robot from any manufacturer in the catalog, then check payload and workspace. Blue shows an estimated reach envelope; orange is your work area. Models without kinematics use a simplified estimate.
Preview: GLB, self-contained glTF, STL, OBJ or USDZ. Have USD, USDA or USDC? See the asset guide below.
Heaviest object with the gripper.
EXPLORE THE CATALOG
Search within the task selected in Project. Select a row to preview a robot, then choose it for your setup. Integration status is separate from payload and reach.
Use your CAD export or NVIDIA SimReady USD assets from the Hub. Keep the original USD files and their dependencies for simulation.
Pick a SimReady object, or export a supported GLB or USDZ. This advisor uses lightweight geometry previews; it does not import an entire USD stage or run its physics. Confirm scale and mass after import.
In your simulator, configure collisions, joints, materials, cameras, actions and the task. An asset alone is not an evaluation environment. EnvHub connects supported environments to LeRobot.
The policy follows from your task. Pick the GPU you have, or train on Hugging Face Jobs and pay by the second.
Where inference happens decides the hardware, the network you need and the deployment technique.
Hardware bought, training and cloud inference, and electricity, in US dollars.
Manufacturing (rough), training, and operation of the robot and its computers, in kg CO2e.
Indicative CO₂ scenario, not a verified product footprint. Electricity factors use Ember’s 2025 generation data. Manufacturing factors remain generic assumptions. Read the sources, exact values and limits →
Your chosen learning path, hardware, assumptions and next steps.
Written by an AI model from your description and your setup. Products, numbers and commands come only from this page.
Run pip install lerobot and hf auth login first. Full walkthrough: .
pip install lerobot
hf auth login