Action Chunking

Action chunking studies how policies can predict and execute sequences of actions together instead of choosing each command separately, reducing inference and decision overhead in tasks such as robotic manipulation and navigation. Methods include Action Chunking Transformers, diffusion- and flow-based vision-language-action policies, and reinforcement-learning approaches that learn chunk-level values or variable execution lengths. A major line of work adapts when to replan using signals such as prediction uncertainty, attention, denoising behavior, task phase, or value estimates; other methods smooth transitions and manage asynchronous inference. Results on benchmarks such as LIBERO, RoboTwin, and OGBench, alongside robot trials, show gains in efficiency and sometimes success, but longer open-loop execution can miss important changes, so balancing temporal consistency against responsive feedback remains central.