ACT from Scratch: Why Robot Policies Predict Action Chunks

ACT from Scratch: Why Robot Policies Predict Action Chunks

Action Chunking Transformer, or ACT, is an imitation learning architecture for robot manipulation.

The key idea is simple:

Instead of predicting only the next action, predict a short sequence of future actions.

This sequence is called an action chunk.

Why single-step behavior cloning can fail

In standard behavior cloning, the policy learns:

\[ \pi_\theta(a_t \mid o_t) \]

Given the current observation \(o_t\), the policy predicts one action \(a_t\).

The problem is that small mistakes accumulate. If the robot makes a slight error, it may enter a state that was rare in the training data. Then the policy becomes less reliable.

This is called compounding error.

Action chunking

ACT predicts:

\[ \pi_\theta(a_{t:t+k-1} \mid o_t) \]

So the model outputs multiple future actions:

\[ (a_t, a_{t+1}, \dots, a_{t+k-1}) \]

This helps because the model learns short-horizon motion structure instead of isolated actions.