SO101-Nexus

SO101-Nexus

Full-stack robot learning for the SO-101 arm: teleoperation, imitation learning, and RL in MuJoCo.

SO101-Nexus takes an SO-101 robot from demonstrations to a trained policy in one installable library. It connects physical leader-arm teleoperation, LeRobot-compatible dataset recording, Gymnasium manipulation environments, and training baselines, all on MuJoCo.

The pipeline is deliberately boring: record demonstrations, clone them with behavior cloning, then reinforce with RL.

Demo

Recorded MuJoCo teleoperation rollouts are published as LeRobot datasets:

What you get

  • Teleoperation. Drive a simulated follower with a physical SO-100 or SO-101 leader arm and save LeRobot v3 datasets, cameras included.
  • Six manipulation tasks. PickLift, PickAndPlace, StackCube, Touch, LookAt, and Move, all on the standard Gymnasium API.
  • Imitation learning and RL. Behavior cloning from your demos, then PPO fine-tuning that keeps anchoring to them.
  • Configurable scenes. Swap objects, add distractors, randomize colors and spawn regions, tune rewards, and choose observation components.
  • LeRobot-native. Datasets, processors, and hardware drivers come from LeRobot rather than a parallel implementation.
  • GPU-parallel Warp backend. The same six tasks as batched vector environments for large-scale RL (optional, experimental, NVIDIA only).

On this page