Gaitor: Learning a Unified Representation for Continuous Gait Transition and Terrain Traversal for Quadruped Robots

Alexander Luis Mitchell, Wolfgang Merkt, Aristotelis Papatheodorou, Ioannis Havoutis, Ingmar Posner

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Keywords: Representation Learning, Learning for Control, Quadruped Robots

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We create a structured latent-space and solve for trajectories within resulting in continuous transition between quadruped gaits and terrain traversal.

Abstract: The current state-of-the-art in quadruped locomotion is able to produce robust motion for terrain traversal but requires the segmentation of a desired trajectory into a discrete set of skills such as trot, crawl and pace. This misses the opportunity to leverage commonalities between individual gait types for efficient learning and are unable to smoothly transition between them. Here we present Gaitor, which creates a learnt representation capturing correlations across multiple distinct gait types resulting in the discovery of smooth transitions between motions. In particular, this representation is compact meaning that information common to all gait types is shared. The emerging structure is interpretable in that it encodes phase correlations between the different gait types which can be leveraged to produce smooth gait transitions. In addition, foot swing characteristics are disentangled and directly addressable. Together with a rudimentary terrain encoding and a learned planner operating in this structured latent representation, Gaitor is able to take motion commands including gait type and characteristics from a user while reacting to uneven terrain. We evaluate Gaitor in both simulated and real-world settings, such as climbing over raised platforms, on an ANYmal C platform. To the best of our knowledge, this is the first work learning an interpretable unified-latent representation for multiple gaits, resulting in smooth and natural looking gait transitions between trot and crawl on a real quadruped robot.

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: 7464