On the Emergence of Whole-body Strategies from Humanoid Robot Push-recovery Learning

TL;DR

This paper presents a model-free deep reinforcement learning approach to develop robust, whole-body push-recovery strategies for humanoid robots, demonstrating effective generalization and robustness in simulation.

Contribution

It introduces a novel RL-based method for high-dimensional humanoid control that outperforms traditional control systems in robustness and generalization.

Findings

Successful learning of multiple push-recovery behaviors

Demonstrated robustness in out-of-sample tasks

Effective policy generalization in simulation

Abstract

Balancing and push-recovery are essential capabilities enabling humanoid robots to solve complex locomotion tasks. In this context, classical control systems tend to be based on simplified physical models and hard-coded strategies. Although successful in specific scenarios, this approach requires demanding tuning of parameters and switching logic between specifically-designed controllers for handling more general perturbations. We apply model-free Deep Reinforcement Learning for training a general and robust humanoid push-recovery policy in a simulation environment. Our method targets high-dimensional whole-body humanoid control and is validated on the iCub humanoid. Reward components incorporating expert knowledge on humanoid control enable fast learning of several robust behaviors by the same policy, spanning the entire body. We validate our method with extensive quantitative analyses in simulation, including out-of-sample tasks which demonstrate policy robustness and generalization, both key requirements towards real-world robot deployment.