Event
PRISM Proposes Polynomial Representation for Interactive Structured Motion Control
1 reports · 1 sources
Seung Hyun Lee and Stella X. Yu propose PRISM, a physics-inspired policy representation method that learns multiplicative interactions among variables by recursively adding scaled bilinear features, without enumerating all polynomial terms. The method only requires replacing the first affine layer of a multilayer perceptron (MLP), and can be used for reinforcement learning and imitation learning without changing training objectives, sensing, or controllers. The paper reports that in humanoid locomotion and contact-intensive manipulation tasks, PRISM raises the survival rate from 51.0% to 92.5%, outperforming larger MLPs; under out-of-distribution changes in friction coefficient, robot mass, and payload, it retains its advantage without retraining. In fault recovery tasks, PRISM raises the success rate from 40.9% to 52.3% without introducing explicit admittance control, and improves manipulation performance on real robots.
Report timeline
- PRISM: Improving Robot Motor Control Policy Learning with Polynomial Representations
PRISM (Polynomial Representations for Interaction Structured Motor Control) is a physics-inspired policy representation method that learns multiplicative interactions among variables by recursively adding scaled bilinear features. It only requires replacing the first affine layer of a multilayer perceptron (MLP) and can be used for reinforcement learning and imitation learning without changing training objectives, sensing, or controllers.
arXiv cs.RO ·