Skip to main content
Once your dataset is uploaded, you can train an ACT policy from the Train tab inside the skill page. ACT (Action Chunking with Transformers) is a visuomotor policy architecture that takes camera images and joint positions as input and predicts a chunk of future actions at once. The “chunking” makes the output temporally smooth and reduces compounding errors compared to single-step prediction.

​
Configure hyperparameters

The Train tab shows your dataset summary and a set of tunable hyperparameters. The defaults work well for most tasks — adjust them only if you have a reason to.
ParameterDefaultWhat it controls
Chunk size30Number of future actions predicted per inference step. Larger values produce smoother but less reactive motion.
Batch size96Training examples per gradient step. Larger batches are more stable but use more GPU memory.
Max steps120,000Total training iterations. More steps can improve quality but eventually overfit on small datasets.
Learning rate5e-5Step size for updating the transformer weights.
LR backbone5e-5Step size for the vision backbone (ResNet18). Lower values fine-tune vision features more gently.
Tap the ? icon next to the hyperparameters for an in-app explanation of each one.

​
When to change the defaults

  • Small dataset (50–80 episodes): Lower max steps to ~80,000 to avoid overfitting.
  • Long episodes or complex task: Increase max steps to 150,000–200,000.
  • Robot seems to hesitate during execution: Try a larger chunk size (50–80) for smoother output.
  • Robot overshoots or ignores corrections: Try a smaller chunk size (15–20) for more reactive behavior.

​
Start a training run

1

Verify sync status

The Train tab shows a dataset card. Confirm the sync badge is green and the episode count looks correct. If it says “Not synced,” go back to the Record tab and upload first.
2

Adjust parameters (optional)

Edit any hyperparameters you want to change, or leave the defaults.
3

Launch training

Tap Start Training Run. Confirm in the dialog. The app creates a run on Innate’s cloud and switches to the Runs tab.
Training runs on Innate’s GPU servers. A typical run with default settings takes 1–3 hours depending on dataset size.
Each robot can have one active training run at a time by default. If you need concurrent runs, reach out on Discord for approval.

​
Monitor a run

The Runs tab shows all active (non-completed) training jobs for this skill. Each run card displays:
  • Run ID — a unique identifier
  • Status — the current stage in the pipeline
  • A progress indicator

​
Training run lifecycle

StatusMeaning
Waiting for approvalRun is queued and pending GPU allocation
ApprovedResources allocated, about to start
BootingTraining instance is spinning up
RunningTraining is in progress
DoneTraining finished, model is ready to download
You can safely close the app or turn off your robot’s screen while training runs. The job continues on the cloud. Status updates resume when you reopen the skill page.

​
What happens during training

Behind the scenes, the training server:
  1. Loads your episodes (images, joint positions, velocities) into a normalized dataset
  2. Trains an ACT model with a ResNet18 vision backbone and a transformer encoder-decoder
  3. Uses a variational autoencoder (VAE) to learn a latent action distribution
  4. Saves checkpoints periodically throughout training
  5. Produces a final checkpoint (.pth) and dataset statistics file (.pt)
The model learns to map what the robot sees and feels to the actions you demonstrated — effectively learning to imitate your behavior.

​
Next steps

When the run status reaches Done, head to the deploy page to download and activate your trained skill.