No-Hardware Quickstart
Complete the record -> clone -> reinforce pipeline using published demonstration datasets, no leader arm required.
You do not need a physical SO-100 or SO-101 leader arm to try the full pipeline. SO101-Nexus publishes LeRobot demonstration datasets, so you can go straight from clone to reinforce and see a trained policy without ever touching hardware.
| Task | Published dataset |
|---|---|
| PickLift | johnsutor/MuJoCoPickLift-v1 |
| PickAndPlace | johnsutor/MuJoCoPickAndPlace-v1 |
The default examples/bc_ppo_warp.py already uses johnsutor/MuJoCoPickLift-v1, so a single
command runs both the clone and the reinforce stages (BC-pretrain, then PPO fine-tuning with
a persistent BC anchor):
uv run --extra warp --extra train python examples/bc_ppo_warp.pyThat is the entire WarpPickLift-v1 pipeline: it downloads the demo dataset, BC-pretrains
the actor, fine-tunes with PPO, and writes runs/WarpPickLift-v1__*/best_agent.pt.
Evaluate the result deterministically:
uv run --extra warp python examples/eval_warp.py \
--env-id WarpPickLift-v1 \
--checkpoint "runs/WarpPickLift-v1__*/best_agent.pt"Prefer a browser? The BC + PPO Colab runs this exact pipeline end to end with embedded TensorBoard and a rollout video. Just pick a GPU runtime and run all cells.
Add your own demonstrations later
When you have a leader arm, the record stage slots in front of this: teleoperate with
so101-nexus teleop, then point the same command at your
dataset with --demo-repo your-user/YourDataset. Everything downstream is unchanged. See the
full Record -> Clone -> Reinforce walkthrough and the
Behavior Cloning details.