ODE Policy: Robot Learning via Action Dynamics, Not Actions
The authors report the highest success rate on 14 real-world robot tasks, with action trajectories continuous by construction.
At deployment, ODE Policy integrates each action from the previous one using a learned vector field. In imitation learning, it reports 5.9 times the success rate of flow matching after one epoch and generates actions 10.8 times faster. The real-world tests span single-arm, dual-arm, dexterous and humanoid robots. In reinforcement learning, it also speeds up SuperMimic’s humanoid motion tracking and smooths its actions.
HF Daily Papers' note
At deployment, ODE Policy integrates each action from the previous one using a learned vector field. In imitation learning, it reports 5.9 times the success rate of flow matching after one epoch and generates actions 10.8 times faster. The real-world tests span single-arm, dual-arm, dexterous and humanoid robots. In reinforcement learning, it also speeds up SuperMimic’s humanoid motion tracking and smooths its actions.
HF Daily Papers' note
score 5