Reproduce “Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries For Zero-Shot Policy Adaptation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiming Xu, Weitao Zhou, Xianghui Pan, Nanshan Deng, Chengju Liu, Qijun Chen, Chenpeng Yao
Latent Dynamics Geometries (LDG) is presented as an outcome-centric framework for zero-shot policy adaptation, using semi-supervised contrastive learning within a variational inference setup to construct a smooth and structured latent space of interaction outcomes. The method enables policies to robustly adapt to unmodeled, time-varying, and even catastrophic dynamics shifts in robotics, outperforming parameter-centric baselines on MuJoCo tasks, especially under structural failures.
Key insights
- The sub-optimality of cross-domain adaptation monotonically increases with the encoder's Lipschitz constant and a local consistency metric, emphasizing the importance of smooth and locally consistent latent representations.
- Contrastive learning, specifically the InfoNCE loss, effectively shapes the latent dynamics geometry by enforcing local consistency, promoting global uniformity, and implicitly filtering nuisance variables irrelevant to dynamics.
- An 'outcome-centric' perspective, which learns from how dynamics affect interaction outcomes, is more robust to unmodeled and catastrophic dynamics shifts than 'parameter-centric' methods that rely on explicit physical parameters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.