SO101-Nexus: full-stack robot learning for the SO-101 arm

Beta: APIs may change between releases. Feedback and bug reports are welcome.

Full-stack robot learning for the SO-101 arm: teleoperation, imitation learning, and RL in MuJoCo. One installable library that takes a robot from demonstrations to a trained policy, built on LeRobot and Gymnasium.

pip install so101-nexus

Full documentation: so101-nexus.com/docs.

The workflow

Record, then clone, then reinforce. Each stage hands one artifact to the next.

1. Record. Drive a simulated follower with a physical SO-100 or SO-101 leader arm and save LeRobot v3 datasets.

uvx --from "so101-nexus[teleop]" so101-nexus teleop --leader-port /dev/ttyACM0

2. Clone. Bootstrap a policy from those demonstrations with behavior cloning.

3. Reinforce. Fine-tune with PPO on the GPU-parallel Warp backend, anchored to the demos.

Stages 2 and 3 are one command. No leader arm? It defaults to a published dataset, so this runs end to end on its own:

uv run --extra warp --extra train python examples/bc_ppo_warp.py

See the workflow walkthrough for the full path.

Run an environment

import gymnasium as gym
import so101_nexus.mujoco  # registers the MuJoCo env ids

env = gym.make("MuJoCoPickLift-v1", render_mode="rgb_array")
obs, info = env.reset()

for _ in range(256):
    obs, reward, terminated, truncated, info = env.step(env.action_space.sample())
    if terminated or truncated:
        obs, info = env.reset()

env.close()

Six SO-101 manipulation tasks ship on MuJoCo: PickLift, PickAndPlace, StackCube, Touch, LookAt, and Move. The optional MuJoCo Warp backend (so101-nexus[warp]) registers the same six as GPU-parallel batched vector environments for large-scale RL. See the environment reference.

The same environments are published as a LeRobot EnvHub package, so a LeRobot user reaches them with one call and no import of their own. The Hub files are shims over this library, so pip install so101-nexus is still the prerequisite:

from lerobot.envs.factory import make_env

envs = make_env(
    "johnsutor/so101-nexus-envs:envs/MuJoCoPickLift-v1.py",
    n_envs=4,
    trust_remote_code=True,
)
env = envs["MuJoCoPickLift-v1"][0]

Why

Plenty of SO-101 tooling exists, but little of it connects teleoperation, LeRobot datasets, simulated environments, and training loops into one workflow. SO101-Nexus is that connection: collect demonstrations, replay and evaluate them in matching SO-101 environments, bootstrap with imitation learning, then fine-tune with RL.

• Teleoperation recorder with a Gradio UI, writing LeRobot v3 datasets with SO follower state and action units plus wrist and overhead camera fields.
• Gymnasium environments with configurable objects, distractors, colors, spawn regions, rewards, and observation components.
• Training baselines for behavior cloning and PPO, plus LeRobot processors and policy adapters for evaluating real policies.
• Optional GPU-parallel Warp backend (experimental, NVIDIA and CUDA only) for batched RL, and an optional ROCm extra for training on AMD hardware.

Recorded MuJoCo teleoperation datasets are published on Hugging Face: MuJoCoPickLift-v1 (viewer), MuJoCoPickAndPlace-v1 (viewer).

Roadmap

• MuJoCo environments for the SO-101 arm
• SO-101 tasks: Touch, LookAt, Move, PickLift, PickAndPlace, StackCube
• Physical leader-arm teleop recorder for LeRobot datasets
• MuJoCo Warp backend for GPU-parallel throughput
• Stronger training baselines and exemplars for every environment
• Integration with the LeRobot Hub

Development

git clone https://github.com/johnsutor/so101-nexus.git
cd so101-nexus
uv sync

make test       # run all tests
make format     # format code
make lint       # lint code

See CONTRIBUTING.md, and Stability and versioning for the public-API and release policy.

License

This repository's source code is available under the Apache-2.0 License.