Direct Policy Optimization with Differentiable Physical Consistency for Dexterous Manipulation

Authors:

Abstract
We present a gradient-based direct policy optimization method for efficiently learning dexterous manipulation tasks in simulation. To overcome local minima and discontinuities, we use a physical consistency loss with additional control variables for contact points and forces. During training, the weights of a neural network are optimized to simultaneously fulfill task goals and to minimize physical inconsistencies. Our method is able to learn policies for manipulating objects with a humanoid hand-arm system within a few minutes on a single CPU.

This publication has 10 references indexed in Scilit: