Skip to main content
Axol records episodes in the LeRobot dataset format, so you train a policy with standard LeRobot training. Record a dataset with Data Collection, point LeRobot’s training script at it, then run the checkpoint on the robot with Run Policy. Any LeRobot training tutorial applies. Use your own dataset’s repo_id in place of the tutorial’s. The links at the end of this page are good places to start.
Train only on datasets you made or got from a source you trust. Dataset loading is not sandboxed.

​
Which policy to train

Start with ACT (Action Chunking with Transformers). It is small (about 80M parameters), trains in a few hours on one GPU, and often works well with about 50 demonstrations. Once you have a working baseline, you can try larger vision-language-action models such as SmolVLA.

​
Train it

Install the lerobot extra on a computer with a CUDA GPU. Then train from your dataset’s Hub repo_id:
Checkpoints are written under outputs/train/.../checkpoints. With --policy.repo_id, the trained policy is also uploaded to the Hub, so the robot and any inference server can download it. Plan for a few hours for about 100k steps on one GPU.
No GPU? The ACT tutorial links a Google Colab notebook for training in the cloud.

​
Mantis datasets: use axol mantis.train

Episodes recorded with the Mantis rigs (axol collect-data --mantis true) store absolute gripper poses. Each take’s frame is set where the operator engaged. A policy trained on these absolute poses learns positions in that one frame, so it doesn’t carry over when the robot starts somewhere else. axol mantis.train fixes this. It runs lerobot-train with the same flags and checkpoint format, and trains the policy to predict each chunk of actions relative to the pose it starts from. The checkpoint stores this setting, so you run it with axol run-policy --policy_path <ckpt> ... as usual.
Plain lerobot-train also runs on a Mantis dataset without errors, but the policy learns fixed positions and doesn’t adapt to a different start pose. Use axol mantis.train for anything recorded with the rigs. You can also use it on Axol data recorded with --robot_config.observe_cartesian true to get the same benefit.
  • Train on your own computer. axol mantis.train refuses remote HF Jobs (--job.target other than local).
  • Chunk blending is approximate. Each relative chunk is tied to the pose it was predicted from. Running each chunk from the queue as predicted is exact. --aggregate_fn temporal_ensemble in run-policy blends overlapping chunks, which works well in practice but is an approximation.
The dataset stays in the standard LeRobot format, so Mantis and on-robot Cartesian episodes can share one dataset.

​
LeRobot tutorials

ACT

The recommended first policy: architecture, training and tips.

Imitation learning walkthrough

Record, train and evaluate with lerobot-train.

SmolVLA

Fine-tune Hugging Face’s small vision-language-action model.

LeRobot docs

Full documentation, models and datasets.

​
Next steps

Run Policy

Run your trained checkpoint on the robot.

Data Collection

Record more episodes to improve the policy.