Action Chunking (ACT)

A learned policy that outputs one action per step can jitter, since each prediction has its own small error, and it can hesitate between two ways of doing a task. The Action Chunking with Transformers policy, ACT, predicts a whole sequence of future actions at once and executes several before predicting again. Blending overlapping chunks, called temporal ensembling, smooths it further.

Policy output Chunk length k Prediction noise

Chunks

At time $t$ the policy predicts the next $k$ actions, $\hat a_{t:t+k} = \pi(o_t)$. Executing the chunk open-loop for $k$ steps reduces how often the policy is queried, so per-query noise has less effect. With temporal ensembling, a new chunk is predicted every step, and the action actually sent is a weighted average of all the chunks that cover the current step:

$$a_t = \frac{\sum_i w_i\,\hat a_t^{(i)}}{\sum_i w_i}, \qquad w_i = e^{-m\,i}$$

What to notice

One action per step gives a jittery path. Chunks are smoother, but a long chunk reacts late, and there can be a jump at each chunk boundary. Ensembling removes the jumps and keeps reacting every step. ACT was introduced with the low-cost ALOHA bimanual system, where these smooth, precise actions made fine tasks like threading and battery insertion possible from about 50 demonstrations.