Record -> Clone -> Reinforce

SO101-Nexus is built around one boring pipeline:

1. Record demonstrations by teleoperating a simulated follower with a physical leader arm, saved as LeRobot datasets.
2. Clone a policy from those demonstrations with behavior cloning (BC).
3. Reinforce the cloned policy with RL fine-tuning (PPO), anchored to the demos.

This page walks the whole path on WarpPickLift-v1, the task with a published, seed-validated demo-seeded recipe. Each stage hands off one artifact to the next:

The pipeline

1. Record demonstrations

Teleoperate a simulated SO-101 follower with a physical SO-100 or SO-101 leader arm. The recorder saves episodes as a LeRobot v3 dataset, ready to push to the Hugging Face Hub.

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

The artifact you carry forward is the dataset repo id (e.g. your-user/MuJoCoPickLift-v1). See Teleoperation for hardware setup, camera fields, and Hub upload.

2. Clone with behavior cloning

examples/bc_ppo_warp.py BC-pretrains the actor mean on your demo transitions before any online training. Point it at your dataset with --demo-repo:

uv run --extra warp --extra train python examples/bc_ppo_warp.py \
    --demo-repo your-user/MuJoCoPickLift-v1 \
    --env-id WarpPickLift-v1

The artifact is a BC-seeded actor: the policy now starts near the demonstrations' actual grasp-lift behavior instead of a random init. See Behavior Cloning for running BC on your own dataset, inspecting the clone, and the action-space details.

3. Reinforce with RL

The same command continues into PPO fine-tuning. With --use-demos (the default), a persistent BC loss (--bc-coef, default 0.1) keeps anchoring the actor toward the demos through training, so the clone is reinforced rather than overwritten. On WarpPickLift-v1 the seed that stalls under vanilla PPO (best_success=0.037) is rescued to best_success=0.993 by demo-seeding alone.

Evaluate the saved checkpoint deterministically:

uv run --extra warp python examples/eval_warp.py \
    --env-id WarpPickLift-v1 \
    --checkpoint "runs/WarpPickLift-v1__*/best_agent.pt"

For the full RL recipe, per-task commands, and tuning notes, see Training with PPO.

The MuJoCo -> Warp handoff

Recording runs on the MuJoCo backend (live viewer, physical leader arm); training runs on the Warp backend (GPU-parallel batched envs). Two things are handled for you, but you should know they exist:

• Action units. Demos store absolute joint-position targets. Training uses pd_joint_delta_pos, so demo actions are recomputed as the per-step delta between consecutive recorded joint states, normalized by the env's _DELTA_ACTION_SCALE. Don't switch control modes between record and train.
• Physics divergence. Warp uses an implicit integrator without the noslip constraint, so a policy may need light re-tuning when moved between backends. See Backend Support for the details.