Direct Policy Optimization with Differentiable Physical Consistency for Dexterous Manipulation
Authors:
- 27 December 2021
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 650-655
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.Keywords
This publication has 10 references indexed in Scilit:
- NeuralSim: Augmenting Differentiable Simulators with Neural Networks2021
- Learning to Solve a Rubik’s Cube with a Dexterous Hand2019
- TaichiACM Transactions on Graphics, 2019
- EA-CG: An Approximate Second-Order Method for Training Fully-Connected Neural NetworksProceedings of the AAAI Conference on Artificial Intelligence, 2019
- ChainQueen: A Real-Time Differentiable Physical Simulator for Soft Robotics2019
- Neural Information Processing2018
- TRAC-IK: An open-source library for improved solving of generic inverse kinematics2015
- Motion planning with sequential convex optimization and convex collision checkingThe International Journal of Robotics Research, 2014
- CHOMP: Covariant Hamiltonian optimization for motion planningThe International Journal of Robotics Research, 2013
- Discovery of complex behaviors through contact-invariant optimizationACM Transactions on Graphics, 2012