How to train your first robot policy

Training sounds like the scary part. It is actually one command and some waiting. The skill is in the data you already recorded, not in the command.

What does "training a policy" actually do?

Your episodes are examples: cameras saw this, motors did that. Training compresses those examples into a model file that, shown a new camera frame, outputs what the motors should do next. Fine-tuning starts from a model that already knows general movement and specializes it to your task.

The first command: ACT

lerobot-train \
  --dataset.repo_id=YOURNAME/first-pick-place \
  --policy.type=act \
  --output_dir=outputs/train/act_first \
  --job_name=act_first \
  --policy.device=cuda \
  --policy.repo_id=YOURNAME/my_first_policy

On Apple silicon use --policy.device=mps. The official docs say plainly: training should take several hours. Start it before dinner, not before a meeting.

No GPU? Two documented routes

The docs link official Colab notebooks for both ACT and SmolVLA, and Hugging Face Jobs runs the same command on rented hardware by adding --job.target=a10g-small (list options and prices with hf jobs hardware). For SmolVLA scale, the documented reference point is about 4 hours for a 20,000-step fine-tune on one A100.

Then watch it move on its own

lerobot-rollout \
  --strategy.type=base \
  --policy.path=YOURNAME/my_first_policy \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM1 \
  --task="Pick up the black cube and place it in the bin" \
  --duration=60

The first time your arm does the task with nobody touching the leader is one of the best moments available in this hobby. Film it. Details and flags in the official training docs and the SmolVLA page.