ODEAction replaces action chunks with continuously evolving robot actions

Researchers led by Qishen Wang, with Cheng Chi among the authors, have described ODEAction in an arXiv paper, a way for vision-language-action models to generate robot actions without splitting them into chunks, which can cause jumps between segments. The model learns a field that describes how the current action should change, and during execution it integrates that field with an ordinary differential equation, starting from the current action.

ODEAction works as an action head with any vision-language backbone. In tests on LIBERO, EBench, RoboTwin and real robots, it improved task performance and the quality of trajectories over matched chunking baselines, especially on fine manipulation tasks, the authors report.