Skip to main content
You teach MARS a task by doing it yourself a few dozen times, with the leader arm or the phone app. A policy learns from those demonstrations. There are two ways to train it, and they end in different places: Both start from the same recordings: a dataset you record in the apps can also go to Hugging Face for LeRobot.
For a motion that never changes, like a wave, record it once as a replay skill instead of training.

​
Train in the Innate cloud

Record a dataset and start a run from the app. Innate trains an ACT policy, and the robot downloads it automatically and runs it as a skill. You don’t need a GPU or any ML setup. The policy looks at both cameras and the arm’s joints, and outputs arm and base commands 25 times a second.
1

Record a dataset

Teleoperate the task 30 or more times. Record a dataset →
2

Train

Launch a run with the default settings. Train in the cloud →
3

Deploy and iterate

Run the skill, see where it fails, add data, retrain. Deploy and iterate →
Every step works in both the phone app and the web app:

​
Train with LeRobot

LeRobot is Hugging Face’s robot learning library. MARS works with its standard commands: you record or upload a dataset, train any of its policies on your own GPU, and run the policy with lerobot-rollout while your computer stays connected to the robot. Train with LeRobot →