Train & Deploy

Choose a Policy Architecture

Step 1: Train with ACT

python -m lerobot.scripts.train \
  --dataset_repo_id=your-username/o6-pick-place \
  --policy.type=act \
  --policy.chunk_size=100 \
  --training.num_epochs=200 \
  --training.batch_size=32 \
  --training.lr=1e-4 \
  --output_dir=./checkpoints/o6-act-v1

# Monitor training loss
tensorboard --logdir=./checkpoints/o6-act-v1/logs/

ACT training typically stabilizes within 100–150 epochs on a 50-episode dataset. Watch for the validation loss to plateau before stopping. Do not stop early based on training time alone.

Step 2: Offline Evaluation

Before deploying to hardware, evaluate the checkpoint on held-out episodes from your dataset:

python -m lerobot.scripts.eval \
  --pretrained_policy_name_or_path=./checkpoints/o6-act-v1 \
  --env.type=linker_bot_o6_sim \
  --eval.n_episodes=20 \
  --eval.batch_size=10

If your environment does not have a sim model, use the dataset replay evaluator instead:

python -m lerobot.scripts.visualize_dataset \
  --repo-id your-username/o6-pick-place \
  --episode-index 0 \
  --policy-checkpoint ./checkpoints/o6-act-v1

Step 3: Live Deployment on the O6

python -m lerobot.scripts.control_robot \
  --robot.type=linker_bot_o6 \
  --control.type=evaluate \
  --pretrained_policy_name_or_path=./checkpoints/o6-act-v1 \
  --control.fps=30 \
  --control.num_episodes=20 \
  --control.episode_time_s=30

For each trial:

1. Place the object in the standard starting position.
2. Confirm the arm is in home position.
3. Start the episode. Do not intervene unless the arm is about to contact itself or the mount.
4. Record success (task completed) or failure, and note the failure mode.

Interpreting Results

The Data Flywheel

Continuous improvement after your first deployment:

1. Identify your top 2–3 failure modes from the evaluation log.
2. Collect 20–30 targeted demonstrations covering those failures.
3. Mix new episodes with your original dataset (50/50 or weight toward failures).
4. Retrain from scratch (not from checkpoint — fresh training avoids catastrophic forgetting on small datasets).
5. Re-evaluate. Repeat until target success rate is reached.

Sharing Your Results

# Push your trained policy to HuggingFace Hub
python -m lerobot.scripts.push_policy_to_hub \
  --pretrained_policy_path=./checkpoints/o6-act-v1 \
  --repo-id=your-username/o6-pick-place-act