Subequivariant Morphology-Behavior Co-Evolution in 3D Environments

Runfa Chen, Yulong Xia, Yunfei Xie, Kangyao Huang, Fuchun Sun, Huaping Liu, Tianyu Shi, Xinpeng Di, Jun Sun, Wenbing Huang

Keywords: Geometric Graph, Reinforcement Learning, Morphology-Behavior Co-Evolution, 3D, Subequivariance

Abstract:

The co-evolution of morphology and behavior in 3D space has garnered considerable interest in the field of embodied intelligence.
While recent studies have highlighted the considerable benefits of geometric symmetry for tasks like learning to locomote, navigate, and explore in dynamic 3D environments, its role within co-evolution setup remains unexplored.
Existing benchmarks encounter several key issues: 1) the task lacks consideration for spatial geometric information; 2) the method lacks geometric symmetry to deal with the complexities in 3D environments.
In this work, we propose a novel setup, named Subequivariant Morphology-Behavior Co-Evolution in 3D Environments (3DS-MB), to address the identified limitations.
To be specific, we propose EquiEvo, which injects geometric symmetry, i.e., subequivariance, to construct dynamic, learnable local reference frames, enabling the joint policy to generalize to diverse task spatial structures, thereby improving co-evolution efficiency.
Then, we evaluate EquiEvo on the proposed environments, where our method consistently and significantly outperforms existing approaches in tasks such as locomotion navigation and adversarial scenarios.
Extensive experiments underscore the importance of subequivariance for the co-evolution of morphology and behavior, effective morphology-task mapping and robust morphology-behavior mapping.

Primary Area: learning on graphs and other geometries & topologies

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: 4336