Artificial intelligence
Action chunking
Definition
Action chunking is the prediction or organization of several future robot actions as one sequence. A policy can execute all or part of a chunk before using new observations to produce another sequence.
Updated
Predict a short sequence together
The ACT paper predicts a sequence of future commands from the current observations rather than predicting only the next command. Grouping commands shortens the number of separate prediction decisions needed to cover a task and helps model temporally correlated demonstrations.
For example, a two-arm robot inserting a battery can predict a short coordinated approach sequence instead of estimating every arm command independently.
Prediction horizon and execution horizon differ
A chunk might contain many future actions while the robot executes only its initial portion. Diffusion Policy combines sequence prediction with receding-horizon control, generating a new plan after fresh observations.
ACT also describes temporal ensembling: overlapping predictions for the same future timestep are combined. This differs from simply smoothing neighboring commands after they have been produced.
Longer chunks trade feedback for continuity
Executing a full chunk without fresh observations delays the response to disturbances. The ACT authors note that a naive chunked implementation can also switch abruptly between observations and produce jerky motion.
Chunking is a representation and execution choice, not a complete learning algorithm. An Action Chunking with Transformers model and a diffusion policy can both use chunks while learning and generating them differently.
Sources
Related terms
Action Chunking with Transformers
Action Chunking with Transformers is an imitation-learning algorithm that predicts sequences of robot actions from observations using a transformer-based conditional variational autoencoder. It is usually abbreviated ACT.
Diffusion policy
A diffusion policy generates robot actions through a learned denoising process conditioned on observations. It commonly predicts an action sequence by progressively refining a noisy candidate rather than predicting one action with a single direct regression.
Action tokenization
Action tokenization converts robot actions or action sequences into discrete symbols that a model can predict and decode into control commands. The tokenizer defines how those symbols represent continuous or discrete robot actions.