Making manipulation faster does not always require a bigger neural network. TOPS changes where flow-based action generation begins and how it progresses, tailoring the source distribution and transport schedule to the task. The network architecture remains unchanged.

On LIBERO-10, one TOPS step achieved 69.67% success against 53.44% for ten standard flow-matching^[1] steps, with roughly one-sixth of the inference time. In a real upper-drawer task, the authors report 80% versus 20%. Both latency and accuracy matter in closed-loop control, although the gains depend on the tested tasks and configurations.