single mention · not yet filed
Chunked action prediction
Chunked action prediction predicts chunks of actions, where each chunk covers multiple time steps, instead of one action at a time.1 A larger chunk lets the robot execute more actions before processing new observations, but both very small and very large chunk sizes degrade performance.2
Trade-off Larger chunks speed up inference by deferring new observations, but reduce the robot's responsiveness to environmental changes.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
Our model predicts chunks of actions, where each chunk consists of time steps
— SmolVLA: A vision-language-action model for affordable and efficient roboticsA larger allows the robot to execute more actions at inference time before needing to process new observations and predict the next chunk. However, Table 12 shows that both very small and very large values of degrade performance
— SmolVLA: A vision-language-action model for affordable and efficient roboticswhile acting the entire chunk speeds up inference, it also reduces the robot’s responsiveness to environmental changes
— SmolVLA: A vision-language-action model for affordable and efficient robotics
Also called predicts chunks of actions.