Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries For Zero-Shot Policy Adaptation

Zhiming Xu, Weitao Zhou, Xianghui Pan, Nanshan Deng, Chengju Liu, Qijun Chen, Chenpeng Yao
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:142962-142988, 2026.

Abstract

Real-world dynamics shifts pose a critical challenge for reinforcement learning, yet prior methods typically rely on encoding explicitly identified physical parameters into a latent context, a rigid parameterization that proves brittle to unmodeled or compound dynamics variations. We instead investigate dynamics adaptation through the lens of latent geometry, and show theoretically that target-domain regret is controlled by the Lipschitz smoothness of a trajectory dynamics encoder. We further prove that this Lipschitz constant can be upper-bounded through optimizing a multi-positive InfoNCE objective, yielding a smooth, task-relevant latent topology without privileged dynamics information. On MuJoCo benchmarks, our method significantly outperforms explicit identification baselines under severe dynamics shifts, including unmodeled structural failures, while simultaneously improving in-distribution stability and latent interpretability. Overall, these results validate that controlling latent smoothness is a principled and scalable mechanism for robust adaptation.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-xu26cq, title = {Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries For Zero-Shot Policy Adaptation}, author = {Xu, Zhiming and Zhou, Weitao and Pan, Xianghui and Deng, Nanshan and Liu, Chengju and Chen, Qijun and Yao, Chenpeng}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {142962--142988}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/xu26cq/xu26cq.pdf}, url = {https://proceedings.mlr.press/v306/xu26cq.html}, abstract = {Real-world dynamics shifts pose a critical challenge for reinforcement learning, yet prior methods typically rely on encoding explicitly identified physical parameters into a latent context, a rigid parameterization that proves brittle to unmodeled or compound dynamics variations. We instead investigate dynamics adaptation through the lens of latent geometry, and show theoretically that target-domain regret is controlled by the Lipschitz smoothness of a trajectory dynamics encoder. We further prove that this Lipschitz constant can be upper-bounded through optimizing a multi-positive InfoNCE objective, yielding a smooth, task-relevant latent topology without privileged dynamics information. On MuJoCo benchmarks, our method significantly outperforms explicit identification baselines under severe dynamics shifts, including unmodeled structural failures, while simultaneously improving in-distribution stability and latent interpretability. Overall, these results validate that controlling latent smoothness is a principled and scalable mechanism for robust adaptation.} }
Endnote
%0 Conference Paper %T Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries For Zero-Shot Policy Adaptation %A Zhiming Xu %A Weitao Zhou %A Xianghui Pan %A Nanshan Deng %A Chengju Liu %A Qijun Chen %A Chenpeng Yao %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-xu26cq %I PMLR %P 142962--142988 %U https://proceedings.mlr.press/v306/xu26cq.html %V 306 %X Real-world dynamics shifts pose a critical challenge for reinforcement learning, yet prior methods typically rely on encoding explicitly identified physical parameters into a latent context, a rigid parameterization that proves brittle to unmodeled or compound dynamics variations. We instead investigate dynamics adaptation through the lens of latent geometry, and show theoretically that target-domain regret is controlled by the Lipschitz smoothness of a trajectory dynamics encoder. We further prove that this Lipschitz constant can be upper-bounded through optimizing a multi-positive InfoNCE objective, yielding a smooth, task-relevant latent topology without privileged dynamics information. On MuJoCo benchmarks, our method significantly outperforms explicit identification baselines under severe dynamics shifts, including unmodeled structural failures, while simultaneously improving in-distribution stability and latent interpretability. Overall, these results validate that controlling latent smoothness is a principled and scalable mechanism for robust adaptation.
APA
Xu, Z., Zhou, W., Pan, X., Deng, N., Liu, C., Chen, Q. & Yao, C.. (2026). Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries For Zero-Shot Policy Adaptation. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:142962-142988 Available from https://proceedings.mlr.press/v306/xu26cq.html.

Related Material