Record -> Clone -> Reinforce

SO101-Nexus is built around one pipeline. Record demonstrations by teleoperating a simulated follower with a physical leader arm, which produces a LeRobot dataset. Clone an actor from that dataset with behavior cloning, which produces a BC-pretrained policy. Reinforce that policy with PPO while a persistent BC loss anchors it to the demos, which produces a checkpoint you can evaluate.

The artifact passed between stages is small and explicit: stage 1 hands over a dataset repo id, stage 2 hands over actor weights, stage 3 hands over best_agent.pt.

Start without hardware

You do not need a leader arm. Two demonstration datasets are published, so you can skip the record stage entirely:

examples/bc_ppo_warp.py already defaults to johnsutor/MuJoCoPickLift-v1 and WarpPickLift-v1, so one command runs both the clone and the reinforce stage:

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

It downloads the demos, BC-pretrains the actor, fine-tunes with PPO, and writes runs/WarpPickLift-v1__*/best_agent.pt. Evaluate that checkpoint deterministically:

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

When you do get a leader arm, the record stage slots in front: teleoperate, then point the same command at your dataset with --demo-repo your-user/YourDataset. Nothing downstream changes.

The MuJoCo to Warp handoff

Recording runs on the MuJoCo backend, because it needs a live viewer and a physical leader arm. Training runs on the Warp backend, because it needs thousands of parallel worlds. Two consequences are handled for you, but you should know they exist.

Action units. Demos store absolute joint-position targets, while training uses pd_joint_delta_pos. Demo actions are relabeled into deltas between consecutive recorded joint states, normalized by the environment's _DELTA_ACTION_SCALE. Do not switch control modes between record and train. See Control Modes.

Physics divergence. Warp uses an implicit integrator with no noslip constraint, so a policy moved between backends may need light re-tuning. See Backends.

Next