Canonical Policy: Learning Canonical 3-D Representation for $\text{SE}(3)$-Equivariant Policy
IEEE Transactions on Robotics · 2026 · 42 · 2325-2344
DOI: 10.1109/tro.2026.3697174 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Robotics, 2026. Verfasst von Zhiyuan Zhang und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 61 Literaturangaben. Gefördert von: National Science Foundation, U.S. Department of Agriculture.
Abstract (Originalsprache)
Visual imitation learning has achieved remarkable progress in robotic manipulation, yet generalization to unseen objects, scene layouts, and camera viewpoints remains a key challenge. Recent advances address this by using 3-D point clouds, which provide geometry-aware, appearance-invariant representations, and by incorporating equivariance into policy architectures to exploit spatial symmetries. However, existing equivariant approaches often lack interpretability and rigor due to unstructured integration of equivariant components. We introduce canonical policy, a principled framework for 3-D equivariant imitation learning that unifies 3-D point cloud observations under a canonical representation. We first establish a theory of 3-D canonical representations, enabling equivariant observation-to-action mappings by grouping both seen and novel point clouds to a canonical representation. We then propose a flexible policy learning pipeline that leverages geometric symmetries from canonical representation and the expressiveness of modern generative models. We validate canonical policy on 12 diverse simulated tasks and 4 real-world manipulation tasks across 16 configurations, involving variations in object color, shape, camera viewpoint, and robot platform. Compared to state-of-the-art imitation learning policies, canonical policy achieves an average improvement of 18.0% in simulation and 39.7% in real-world experiments, demonstrating superior generalization capability and sample efficiency.
Förderung
- National Science Foundation
- U.S. Department of Agriculture
Zitieren
APA
Zhang, Z., Xu, Z., Lakamsani, J. N., & She, Y. (2026). Canonical Policy: Learning Canonical 3-D Representation for $\text{SE}(3)$-Equivariant Policy. IEEE Transactions on Robotics, 42, 2325-2344. https://doi.org/10.1109/tro.2026.3697174
Vancouver
Zhang Z, Xu Z, Lakamsani JN, She Y. Canonical Policy: Learning Canonical 3-D Representation for $\text{SE}(3)$-Equivariant Policy. IEEE Transactions on Robotics. 2026;42:2325-2344. doi:10.1109/tro.2026.3697174
BibTeX
@article{Zhang2026,
author = {Zhang, Zhiyuan and Xu, Zhengtong and Lakamsani, Jai Nanda and She, Yu},
title = {Canonical Policy: Learning Canonical 3-D Representation for $\textSE(3)$-Equivariant Policy},
journal = {IEEE Transactions on Robotics},
year = {2026},
volume = {42},
pages = {2325--2344},
doi = {10.1109/tro.2026.3697174}
}
EUVIMED