Skip to main content
Axol records every episode in the LeRobot dataset format, so training a policy is just standard LeRobot training — there’s nothing Axol-specific about it. Once you’ve recorded a dataset with Data Collection, you point LeRobot’s training script at it, and the resulting checkpoint runs back on the robot with Run Policy.
Because the data is plain LeRobot format, any LeRobot training tutorial applies directly — just substitute your own dataset repo_id. The links below are the best places to start.
Train only from datasets you created or obtained from a trusted source. Upstream dataset parsing is not a sandbox for arbitrary Hub repositories or directories supplied by someone else.

​
Which policy to train

ACT (Action Chunking with Transformers) is the recommended starting point: it’s lightweight (~80M params), trains in a few hours on a single GPU, and often reaches high success rates with as few as ~50 demonstrations. Once you have a baseline working, you can explore larger vision-language-action models like SmolVLA.

​
Train it

With the lerobot extra installed on a machine with a CUDA-capable GPU, train directly from your dataset’s Hub repo_id:
Training writes checkpoints under outputs/train/.../checkpoints, and (with --policy.repo_id) pushes the trained policy to the Hub so it’s reachable from the robot machine and any inference server. Expect a few hours for ~100k steps on a single GPU.
No local GPU? The ACT tutorial links a Google Colab notebook you can use to train in the cloud.

​
Mantis datasets: use axol mantis.train

Episodes recorded with the Mantis handheld rigs (axol collect-data --mantis true) store absolute base-frame end-effector poses in the Cartesian schema. Each take is anchored to wherever the demonstrator engaged, so a policy trained on absolute targets learns positions in that one world frame and does not transfer when the robot starts somewhere else. axol mantis.train fixes that at training time. It wraps lerobot-train — same CLI surface (draccus dotted overrides, --config_path, wandb, resume), same checkpoint format — and installs a chunk-relative end-effector processor in the policy’s pipeline: every predicted action chunk is expressed relative to the observation it was predicted from, and the action/state normalisation statistics are recomputed over the relativised values. The relativisation lives in the checkpoint’s processor pipeline, so deployment is unchanged: axol run-policy --policy_path <ckpt> ... as usual.
Plain lerobot-train also runs on a Mantis dataset without error, but it produces a world-anchored policy with no start-pose invariance. Use axol mantis.train for anything recorded with the rigs (or with --robot_config.observe_cartesian true on the robot, if you want the same relative-action generalisation).
Two things to know:
  • Local only. Remote HF Jobs (--job.target other than local) are rejected: the remote pod runs stock lerobot-train and cannot carry the processor patch. Run the command on the training machine.
  • Chunk execution. With relative actions, chunks predicted from different observations are anchored to different reference poses. The default chunk-queue execution in run-policy is exact; --aggregate_fn temporal_ensemble blends near-identical absolute actions from overlapping chunks and works well in practice, but the blend is an approximation.
The dataset itself stays standard LeRobot format, so rig-collected and on-robot Cartesian episodes mix freely in one dataset.

​
LeRobotTutorials

ACT

The recommended first policy — architecture, training, and tips.

Imitation learning walkthrough

End-to-end record → train → evaluate with lerobot-train.

SmolVLA

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

LeRobot docs

Full documentation, models, and datasets.

​
Next steps

Run Policy

Deploy your trained checkpoint on the robot.

Data Collection

Record more episodes to improve the policy.