Code is coming!

The paper is under review.

Abstract

Modular robots have the potential for an unmatched ability to perform versatile and robust locomotion. However, designing effective and adaptive locomotion controllers for modular robots is challenging, resulting in solutions that typically require various forms of prior knowledge.

In this paper, we propose a novel two-level hierarchical locomotion framework for modular quadrupedal robots with a limited amount of prior knowledge. Our approach combines a low-level central pattern generator (CPG)-based controller with a high-level neural network to learn a variety of locomotion tasks using deep reinforcement learning. The low-level CPG controller is preoptimized to generate stable rhythmic walking gaits, while the high-level network is trained to modulate the CPG parameters for achieving the task goal based on high-dimensional inputs, including the states of the robot and user commands. The proposed approach is employed on a simulated modular quadruped. We empirically demonstrate the learned policies using our approach can allow the robot to perform multiple locomotion skills. The results show that our framework outperforms prior model-based and model-free methods in terms of robustness as well as sample efficiency.

Video

Contact

If you have any questions, please feel free to contact Jerry Wang at wjywangjiay_at_gmail.com

March 1, 2021

Posted by Jerry Wang