A Caltech Library Service

Online Learning of Unknown Dynamics for Model-Based Controllers in Legged Locomotion

Sun, Yu and Ubellacker, Wyatt L. and Ma, Wen-Loong and Zhang, Xiang and Wang, Changhao and Csomay-Shanklin, Noel V. and Tomizuka, Masayoshi and Sreenath, Koushil and Ames, Aaron D. (2021) Online Learning of Unknown Dynamics for Model-Based Controllers in Legged Locomotion. IEEE Robotics and Automation Letters, 6 (4). pp. 8442-8449. ISSN 2377-3766. doi:10.1109/lra.2021.3108510.

Full text is not posted in this repository. Consult Related URLs below.

Use this Persistent URL to link to this item:


The performance of a model-based controller can severely suffer when its model inaccurately represents the real world dynamics. We propose to learn a time-varying, locally linear residual model along the robot’s current trajectory, to compensate for the prediction errors of the controller’s model. Supervised learning is performed online, as the robot is running in the unknown environment, using data collected from its immediate past. We theoretically investigate our method in its general formulation, then apply it to a bipedal controller derived from the full-order dynamics of virtual constraints, and a quadrupedal controller derived from a simplified model of contact forces. For a biped in simulation, our method consistently outperforms the baseline and a recent learning-based method. We also experiment with a 12 kg quadruped in simulation and real world, where the baseline fails to walk with 10 kg of payload but our method succeeds.

Item Type:Article
Related URLs:
URLURL TypeDescription
Sun, Yu0000-0001-7626-8191
Ma, Wen-Loong0000-0002-0115-5632
Zhang, Xiang0000-0001-7900-5127
Wang, Changhao0000-0002-5753-1991
Csomay-Shanklin, Noel V.0000-0002-2361-1694
Tomizuka, Masayoshi0000-0003-0206-6639
Sreenath, Koushil0000-0002-5346-3637
Ames, Aaron D.0000-0003-0848-3177
Additional Information:© 2021 IEEE. Manuscript received February 24, 2021; accepted June 13, 2021. Date of publication August 30, 2021; date of current version September 13, 2021. This letter was recommended for publication by Editor Dana Kulic upon evaluation of the Associate Editor and reviewers’ comments. This work was supported in part by DOW Chemical Project 227027AT, in part by NSF NRI Award 1924526, in part by NSF DCSDVTECH Award 1923239, and in part by NSF under Grant CMMI-1944722. We thank Fernando Castañeda for providing his code of the method in [33] to compare with our method. Yu Sun would like to thank his advisors, Alexei A. Efros and Moritz Hardt, for their unwavering support, and Armin Askari, Zihao Chen, Ashish Kumar, John Miller, and Haozhi Qi, for their help.
Funding AgencyGrant Number
Dow Chemical Company227027AT
Subject Keywords:Model learning for control, legged robots
Issue or Number:4
Record Number:CaltechAUTHORS:20211021-172748746
Persistent URL:
Official Citation:Y. Sun et al., "Online Learning of Unknown Dynamics for Model-Based Controllers in Legged Locomotion," in IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 8442-8449, Oct. 2021, doi: 10.1109/LRA.2021.3108510
Usage Policy:No commercial reproduction, distribution, display or performance rights in this work are provided.
ID Code:111574
Deposited By: Tony Diaz
Deposited On:22 Oct 2021 23:03
Last Modified:26 Oct 2021 19:13

Repository Staff Only: item control page