ODEAction: From Predicting Chunks to Evolving Actions

This paper introduces ODEAction, a new paradigm for vision-language-action models that eliminates discontinuities by using ordinary differential equations to continuously evolve actions rather than predicting discrete chunks. The method defines an Action Transition Field to characterize action evolution over time and solves the ODE during inference to generate subsequent actions recursively. Implemented as a modular component compatible with various backbones, ODEAction demonstrates improved task performance and trajectory quality compared to chunk-based baselines, particularly in fine manipulation tasks, across benchmarks including LIBERO, EBench, RoboTwin, and real-robot experiments.