Action Chunking: Multi-Step Action Prediction for Smoother Robot Control

Action chunking is a paradigm for robot policy prediction where the model outputs a sequence of K future actions at each decision point, rather than a single action. The robot then executes some or all of the predicted actions before querying the policy again for a new chunk. This approach was popularized by the ACT model (Action Chunking with Transformers) from Zhao et al. (2023) and is used by Diffusion Policy, pi-zero, and most modern manipulation policy architectures.

The theoretical motivation is to address compounding errors in behavioral cloning. When a policy predicts one action at a time, small prediction errors cause the robot to visit states slightly outside the training distribution. At these unfamiliar states, the next prediction is less accurate, creating a cascade where errors accumulate quadratically with the time horizon. Action chunking breaks this cascade by having the policy commit to a coherent multi-step plan: even if the starting state is slightly off-distribution, the predicted action sequence remains internally consistent, and the robot follows a smooth trajectory rather than accumulating independent errors.

Architecturally, action chunking changes the output layer of the policy from a single action vector to a matrix of shape (K, action_dim), where K is the chunk size. A policy with a 7-dimensional action space (x, y, z, roll, pitch, yaw, gripper) and chunk size 16 outputs 16 x 7 = 112 values at each prediction. The policy must learn temporal structure within the chunk — that the first few actions should approach the object, the middle actions should close the gripper, and the final actions should lift.

The execution strategy determines how chunks interact with the environment. In the simplest approach, the full chunk of K actions is executed open-loop, then a new chunk is predicted. This is efficient (the policy runs at 1/K the decision frequency) but cannot react to disturbances within a chunk. The receding-horizon approach executes only the first N < K actions from each chunk, then re-predicts a new chunk from the current observation. This provides reactivity (re-planning every N steps) while maintaining temporal consistency (each chunk is internally coherent). Most practical systems use N = K/4 to K/2.

Data collection for action chunking models requires attention to temporal consistency and recording frequency that goes beyond standard behavioral cloning data.

The recording frequency must match or exceed the control frequency. If the policy will operate at 20 Hz with 16-step chunks, the data must be recorded at 20 Hz or higher. Down-sampling from higher frequencies is acceptable (record at 50 Hz, down-sample to 20 Hz for training), but up-sampling from lower frequencies introduces interpolation artifacts. Record at the highest practical frequency and down-sample during data loading.

Demonstration quality requirements are stricter for chunking models because the policy must learn temporal patterns within chunks. Demonstrations with frequent stops, reversals, or hesitations create inconsistent chunk targets — the model cannot learn a coherent multi-step plan from data where the demonstrator constantly changes direction. Operators should be trained to execute smooth, committed motions. Filtering demonstrations by trajectory smoothness metrics (jerk, velocity variance) is particularly important for chunking architectures.

Chunk size is a critical hyperparameter. Too small (2-4 steps) and the chunking provides little benefit over single-step prediction. Too large (100+ steps at low control frequencies) and the policy cannot react to disturbances within a chunk. The optimal chunk size correlates with the typical duration of coherent sub-actions in the task. For tabletop pick-and-place at 10 Hz, 16-step chunks (1.6 seconds) cover the approach-grasp-lift sequence well. For longer tasks, shorter chunks with receding-horizon execution provide the right balance of coherence and reactivity.

The data format must preserve exact temporal ordering and synchronization within episodes. Action chunks are extracted from consecutive timesteps in the training data, so any temporal misalignment between observation and action timestamps corrupts the chunk targets. This is more damaging than for single-step BC because the error propagates across the entire chunk length.

Claru's teleoperation data is collected with the temporal quality standards that action chunking architectures demand. Our operators are trained to produce smooth, committed motions without unnecessary hesitations, and our recording pipelines capture all sensor streams at 30-50 Hz with verified temporal synchronization.

For teams training ACT, Diffusion Policy, or VLA models with action chunking, Claru delivers demonstrations where the temporal structure within every potential chunk is clean — approach, grasp, and manipulation phases flow naturally without the stops and restarts that corrupt chunk-level training targets.