ODE Policy learns action dynamics, not actions: 5.9x success over flow matching

micoolcho · x · 2026-11-05

Jianfei Yang's lab (NTU) released ODE Policy (arXiv:2611.01729): instead of predicting discrete actions, a neural ODE learns the vector field of action dynamics, integrating each action from the previous one so trajectories are continuous by construction.

Key results

In imitation learning vs flow matching: 5.9x success rate, 5.4x generalization, 3.9x mode balance after one epoch, and 10.8x faster action generation; executed trajectories are often smoother than human demonstrations
Works as a visuomotor policy, as the action head of VLAs and world action models, and as the actor in model-based RL
Evaluated against 14 baselines on unimanual, bimanual, dexterous, locomotion, and loco-manipulation tasks
Highest success rate on 14 real-world tasks across single-arm, dual-arm, dexterous, and humanoid robots; a lightweight diversity head captures multi-modal behavior by randomizing field direction

Original post →

ODE Policy learns action dynamics, not actions: 5.9x success over flow matching

More from Embodied