Open Collections

Xie, Zhaoming

University of British Columbia

Deep reinforcement learning (DRL) offers a promising approach for the synthesis of control policies for legged robots locomotion. However, it remains challenging to learn policies that are robust to uncertainty in the real world to put on physical robots or policies that can handle complicated environments. In this thesis, we take several significant steps towards efficiently learning legged locomotion skills with DRL. First, we present a framework to learn feedback policies for a bipedal robotCassie, utilizing rough motion sketches. An iterative design process is then proposed to refine, compress and combine policies for effective sim-to-real transfer. Second, we explore the role of dynamics randomization on a quadrupedal robotLaikago. We demonstrate that with appropriate design choices, dynamics randomization is often not necessary for sim-to-real. We further analyze situations that randomization would become necessary. Third, we propose and analyze multiple curriculum learning approaches to solve the challenging stepping stone tasks for bipedal locomotion. We demonstrate that gradually increasing task difficulties can reliably train policies that solve challenging stepping stone sequences. Finally, we investigate the combination of reinforcement learning and model-based control by training quadrupedal policies using a centroidal model.

Text

2021-12-07

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/80396

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/