ODE Policy learns robot action dynamics rather than discrete actions
Original title: ODE Policy: Robot Learning via Action Dynamics, Not Actions
After one epoch it reports 5.9 times flow matching's success and the highest rate on 14 real-world robot tasks.
The paper proposes ODE policy, a neural ordinary differential equation that learns an action vector field rather than discrete action values. Integrated actions are matched to targets over short chunks in imitation learning and single steps in reinforcement learning. A diversity head randomizes field direction for multi-modal behavior. Authors report gains versus 14 baselines and the highest success rate on 14 real-world tasks.
Unconfirmed: Performance figures are reported by the paper authors and are not independently confirmed here.
- In imitation learning it reports 5.4 times generalization and 3.9 times mode balance after one epoch.
- It generates actions 10.8 times faster than flow matching and mostly smoother than human demonstrations.
- Evaluated as a visuomotor policy, VLA and world-action-model head, and model-based RL actor.
- In RL it speeds SuperMimic humanoid tracking and smooths its actions.
- AI models
- ODE PolicySuperMimic
From the original article
1 min read at huggingface.co
Most robot policies, whether trained by imitation learning (IL) or reinforcement learning (RL), map observations to action values at discrete control steps and leave the continuity of physical motion to be learned implicitly. We propose ODE policy, which learns the dynamics of the action rather than the action itself.
By Jindou Jia, Jingliang Li, Kuangji Zuo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.