Skip to main content
You teleoperate the robot through the task while it records both cameras and the arm’s joints at 30 Hz. Each recording is an episode. You need a leader arm. Train on your dataset, then share it on Hugging Face so others can see what MARS can do, reuse your episodes, and train their own policies. The web app's Collect page

​
Record episodes

1

Create the skill

On the Collect page, open the skill picker, choose + New skill… → Learned skill, and name it.
2

Engage the leader arm

Plug the leader arm into your computer and click Connect arm, then Engage follow. Use Chrome or Edge.
3

Record

Press Record, do the task, press Stop, then Save or Discard. Repeat.
The head stays locked while recording, so every episode has the same view.

​
Record good data

The policy copies exactly what you do, mistakes included. How many? At least 30 clean episodes for a simple task, 50–80 to be comfortable, and 150 or more for varied positions, objects, or lighting. 30 clean episodes beat 80 sloppy ones. If the task involves driving, turn on base movement in the recorder. The episode then also records velocity commands and odometry.

​
Review and upload for training

Replay your episodes and delete bad ones: on the web app’s Datasets page, or the skill’s Record tab in the phone app. New episodes show preparing while they encode. To upload for cloud training, tap Upload in the phone app’s Record tab and wait for the green sync badge. The web app uploads automatically when you start training. Only new episodes are sent.

​
Share on Hugging Face

Hugging Face Hugging Face Publish your dataset to the Hugging Face Hub as a standard LeRobot dataset. Anyone can then browse your episodes in LeRobot’s dataset visualizer, and train on them with LeRobot. For an example, see innate-inc/mars-pick-tv.
1

Add your token

Create a token with write permission at huggingface.co/settings/tokens. In the web app, save it under Settings → Keys → Hugging Face.
2

Publish

On the Datasets page, open your dataset and press Publish to Hugging Face. The first time, the robot downloads the LeRobot environment, about 1.5 GB.
3

Name it

Pick your account or an organization, and a repository name. Private is ticked by default: untick it to share the dataset with everyone. Only public datasets open in the visualizer.
Publishing runs in the background, so you can close the dialog and reopen it to check progress. After you record more episodes, publish again: only the new ones are uploaded. Deleted episodes, and ones marked as failed, are left out of the next publish. Datasets you recorded before Innate OS 0.8.0 publish too. Next: Train in the cloud, or train with LeRobot.