Train a Policy - Innate Docs

Start a run

Web app
Phone app

On the Training page, click + Train a skill, pick the skill, and start. The dataset uploads first.

A run takes 1–3 hours. You can close the app or turn off the robot meanwhile. Each robot runs one training job at a time; ask on Discord if you need more.

The web app's Train skill dialog

Settings

The defaults work for most tasks. Change them only to fix a specific problem.

Setting	Default	Meaning
Chunk size	30	Actions predicted at once. Higher is smoother, lower is more reactive.
Batch size	96	Examples per training step
Max steps	120,000	Training length
Learning rate	5e-5	Step size for the transformer
LR backbone	5e-5	Step size for the ResNet18 vision backbone

If	Try
Your dataset is small (50–80 episodes)	Max steps around 80,000, to avoid overfitting
The task is long or complex	Max steps 150,000–200,000
The robot hesitates	Chunk size 50–80
The robot overshoots or ignores corrections	Chunk size 15–20

Monitor

The web app’s Running now card shows progress and an ETA. Click a run for live logs and a Weights & Biases link. The phone app’s Runs tab shows status.

Status	Meaning
Waiting for approval	Queued for a GPU
Approved, Booting	Starting up
Running	Training
Done	Ready to deploy