Interplay Between Task Learning and Skill Discovery for Agile Locomotion

Seungeun Rho, Kartik Garg, Harshil Vagadia, Sehoon Ha

Keywords: Unsupervised Skill Discovery, Reinforcement Learning, Robot Learning, Locomotion

Abstract: Agile locomotion of legged robots, characterized by high momentum and frequent contact changes, is a challenging task that demands precise motor control. Therefore, the training process for such skills often relies on additional techniques, such as reward engineering, expert demonstrations, and curriculum learning. However, these requirements hinder the generalizability of methods because we may lack sufficient prior knowledge or demonstration datasets for some tasks. In this work, we consider the problem of automated learning agile motions using its intrinsic motivation, which can greatly reduce the effort of a human engineer. Inspired by unsupervised skill discovery, our learning framework encourages the agent to explore various skills to maximize the given task reward. Finally, we train a parameter to balance the two distinct rewards through a bi-level optimization process. We demonstrate that our method can train quadrupeds to perform highly agile motions, ranging from crawling, jumping, and leaping to complex maneuvers such as jumping off a perpendicular wall.

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7743