Beta: APIs may change between releases. Feedback and bug reports are welcome.
Record demonstrations, train a policy with behavior cloning, then fine-tune it with reinforcement learning. SO101-Nexus connects SO-100 and SO-101 leader arms, LeRobot datasets, and Gymnasium environments in one Python library.
| I want to... | Go to |
|---|---|
| Install the library | Installation |
| Run a simulated robot | Quickstart |
| Record demonstrations with a leader arm | Teleoperation |
| Train with published demonstrations | Training |
| Choose a task | Environments |
| Use LeRobot EnvHub | LeRobot compatibility |
No leader arm? Start with the published demonstrations in the BC + PPO Colab notebook. Select a GPU runtime and run all cells.
- Record. Control a simulated follower with a physical leader arm and save a LeRobot dataset.
- Clone. Train a policy to reproduce the demonstrations.
- Reinforce. Fine-tune the policy with PPO on the GPU-parallel MuJoCo Warp backend.
The workflow guide connects these stages. The examples index lists the training scripts and notebooks.
MuJoCo provides the default simulation backend. The optional Warp backend supports batched GPU training. See Backends for rendering, hardware requirements, and physics differences.
Follow the source installation guide, then run:
make format lint typecheck
make testSee CONTRIBUTING.md for contribution instructions and Stability and versioning for the release policy.
This repository's source code is available under the Apache-2.0 License.