Robot policies learn action dynamics instead of actions: ODE Policy hits 79.2% on LIBERO after one epoch, versus 43.2% at best for baselines
Synopsis
Most robot policies output action values at discrete control steps and leave the continuity of physical motion to be learned implicitly from data. ODE Policy instead learns a neural ordinary differential equation for the vector field of the action, and at deployment integrates each action from the previous one, so the executed trajectory is continuous by construction. After one epoch on LIBERO it reaches 79.2% success against at most 43.2% for the baselines, and it keeps that early lead as the action head of a VLA and of ImageWAM and as the actor of SuperMimic.
Interpretation
A robot policy no longer emits action values step by step; it emits the rate at which the action changes, and integration turns that into an action sequence. Earlier imitation-learning and reinforcement-learning policies mapped observations directly to action values at discrete control steps, so continuity of motion had to be learned implicitly from data and was not even guaranteed at chunk boundaries. After one epoch on LIBERO, ODE Policy succeeds in 79.2% of episodes against at most 43.2% for the six visuomotor baselines; it passes 85% at epoch 2, an order of magnitude earlier than flow and diffusion policies.
The mechanism is a neural network that serves as the vector field of the action, taking the current action and proprioception and returning its instantaneous velocity. Training integrates over action chunks with a numerical solver and matches only the integrated path to the demonstration, without regressing velocity labels, so gradients pass through a few solver steps. One model is trained on the four LIBERO suites, each task is evaluated with 20 episodes from the benchmark's fixed initial states, and each visuomotor experiment is repeated with three random seeds, reporting mean and standard deviation.
At deployment every action is integrated from the previous one, and successive chunks join end to end, so execution need not wait for a whole chunk. Direct-action policies jump at chunk boundaries, and generative policies need multiple network evaluations per chunk. With the Euler solver, a chunk of 8 actions takes 5.7 ms, 10.8 times faster than FM-DiT; on real robots it is 3.5 to 27 times faster than the diffusion and flow matching baselines.
Perspective
The demonstrated scope covers LIBERO and LIBERO-plus simulation, a 2D navigation task, 14 real-world tasks on a Franka Research 3 arm, a Galaxea R1 Lite dual-arm robot, a Sharpa North dexterous hand and a Unitree G1 humanoid, plus 40 LAFAN1 motion clips tracked by SuperMimic; real-robot methods train for 300 epochs (50 on the single arm, 90 on the humanoid) and are evaluated with 20 trials each.
Multi-modal behavior is verified only on a 2D navigation task and one real bimanual drawer task, and the authors note that such strong multi-modality occurs in only a small part of robot tasks; integrating the gripper as a continuous state lowers success, and early model-free RL experiments showed little gain and sometimes harm.