Models & learning
Action chunking
Action chunking is a policy design in which one observation is used to predict a sequence of several future robot actions rather than a single next action. The chunk can capture short-term motion structure and reduce the effective number of sequential predictions, while its execution and replanning schedule determine how quickly the robot can react to new observations.
Also known as: action chunk, action chunks, temporal action chunking
Updated
Predicting a sequence at once
A single-step policy repeatedly predicts action[t] from the latest observation. An action-chunk policy predicts actions[t:t+H], where H is the action horizon. Action Chunking with Transformers showed that increasing the chunk length substantially improved the reported imitation-learning tasks up to a task-dependent point.
The chunk gives the model a direct representation of short motion sequences. It can reduce compounding error because fewer separate policy calls are needed to cover the same physical interval.
Chunk length and replanning
Executing a whole chunk without observing again is open-loop during that interval. A long chunk can therefore become stale when an object moves or the robot drifts. A short chunk is more reactive but demands more frequent inference.
Policies can replan before a chunk is exhausted. ACT predicts overlapping chunks and combines actions assigned to the same future timestep. Diffusion Policy uses an observation horizon and a receding action horizon. Asynchronous inference can request the next chunk while the robot is still consuming the current one.
Dataset requirements
Training needs consecutive future actions for each sampled observation, accurate timestamps and an episode-boundary mask. Padding must not be treated as a real target. The dataset also needs a declared control frequency because a chunk of 50 actions represents one second at 50 Hz but five seconds at 10 Hz.
Action chunking is independent of the model family. Transformers, diffusion models and flow-matching policies can all predict chunks, and the term does not imply one specific architecture.
Sources
Related terms
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.
Models & learning
Behaviour cloning
Behaviour cloning is a form of imitation learning that fits a policy to expert observation–action pairs as a supervised prediction problem. For humanoid robots, the training examples typically align camera or proprioceptive observations with commands recorded during demonstrations, so the learned policy can reproduce similar behaviour without an explicit reward model.
Models & learning
Diffusion policy
A diffusion policy is a learned robot policy that generates an action or action sequence through iterative denoising conditioned on observations and, where applicable, a task goal. This formulation can represent several plausible behaviours for the same scene and is commonly trained by imitation on time-aligned trajectories of observations and expert actions.
Data & collection
Trajectory
A trajectory is a time-ordered sequence of states or observations, actions and, where applicable, rewards generated as an agent or robot evolves. A complete episode or policy rollout often yields a trajectory, but the terms are not universally identical: trajectories may be partial, while episodes have dataset- or environment-defined boundaries.
Models & learning
Asynchronous inference
Asynchronous inference is a deployment pattern that overlaps robot action execution with prediction of a later action chunk. A robot client consumes queued actions while a policy process or server evaluates a newer observation, then merges or replaces the remaining queue when the new chunk arrives. The pattern reduces idle inference gaps but does not remove latency, stale observations or safety constraints.