MorAL: Learning Morphologically Adaptive Locomotion Controller for Quadrupedal Robots on Challenging Terrains

A Learning-based Morphologically Adaptive Locomotion Controller for legged robots


morphology


Abstract

Due to the rapid development of the quadruped robot industry in the past decade, various commercial quadruped robots have emerged with distinct physical attributes. Different from the previous work in which the designed controller is robot-specific, this article proposes a learning-based control framework -- MorAL, which is adaptive to different morphologies of quadruped robots and challenging terrains. Our framework concurrently trains the control policy and an adaptive module, which considers the temporal robot states. This module empowers the control policy to implicitly online identify different robot platforms' properties and estimate body velocity. Extensive experiments in the real world and simulation demonstrate that our controller enables robots with significantly different morphology to overcome various indoor and outdoor harsh terrains.


Comparison

The comparative evaluation is performed among the following locomotion controllers that only use proprioception, including Vanilla PPO, RMA, Concurrent and DreamWaQ. The following figure illustrates the learning performance of five different controllers in terms of the average rewards. It indicates that MorAL is the most efficient controller in this multi-morphology task.

morphology

To make the comparison more fair, further analysis conducted on their performance in terms of the robot-specific tasks. The following results indicate MorAL still outperforms other baseline methods. The analysis is conduted on four different specific robot configurations, including Aliengo (FE), ANYMal-B (FKBE), Mini Cheetah-FK (FK), and Mini Cheetah-FEBK (FEBK).

morphology

The MorAL algorithm also achieve the smallest velocity estimation error and the most robust velcocity tracking performance.

morphology

Citation

Please cite the following publication when using this benchmark in an academic context:

  1. RA-L
    MorAL: Learning Morphologically Adaptive Locomotion Controller for Quadrupedal Robots on Challenging Terrains
    Z. Luo, Y. Dong, X. Li, R. Huang, Z. Shu, E. Xiao, and P. Lu
    IEEE Robotics and Automation Letters 2024

Acknowledgement

This work was supported by the General Research Fund under Grant 17204222, and in part by the Seed Fund for Collaborative Research and General Funding Scheme-HKU-TCL Joint Research Center for Artificial Intelligence.