RL.md

1. Collect Demonstrations (Teleoperation)

First, run lekiwi_host on the Jetson, then record episodes from the laptop:

Jetson: start the robot server

python -m lerobot.robots.lekiwi.lekiwi_host

Laptop: record teleoperation episodes into a dataset

python -m lerobot record
--robot.type=lekiwi
--dataset.repo_id=/
--dataset.num_episodes=50

You physically teleoperate the leader arm while the system records observations (camera frames, joint positions) and actions.

2. Train a Policy

lerobot supports several imitation learning policies. The two main ones:

ACT (Action Chunking with Transformers) — good default

python -m lerobot train
--policy.type=act
--dataset.repo_id=/

Or Diffusion Policy

python -m lerobot train
--policy.type=diffusion
--dataset.repo_id=/

You already have wandb in your dependencies for tracking training runs. Your pyproject.toml also has the smolvla optional — that's a Vision-Language-Action model if you want to try that route later.

3. Evaluate

Run the trained policy on the real robot:

python -m lerobot eval
--policy.path=outputs/train//checkpoints/last/pretrained_model
--robot.type=lekiwi