Observe Change, Preserve Constancy: DTS-Reg for Robust Genmount World Models under Visual Domain Shifts

【Method & Protocol Preprint — NO measured ID/OOD returns are claimed.】 Robotic vision in the real world is dominated by appearance change—illumination, texture, background, camera pose, lens contamination, and sensor noise—while task-relevant dynamics and reward structure often remain comparatively stable. We study how to train Genmount world models from videos (pixel sequences) so that latent imagination remains usable for end-to-end visuomotor control under visual domain shifts. Inspired by the Eastern philosophical principle “Observe change, preserve constancy” (观其变,守其常), we argue that robustness should not enforce pointwise latent alignment across domains. Instead, it should preserve local relational structure that supports prediction and planning. We propose Dynamic Topological Stability (DTS), a metric that quantifies whether neighborhood relations in latent space remain stable across domains, both at posterior encoding and along multi-step prior rollouts. Building on DTS, we introduce DTS-Reg, a lightweight regularizer that aligns cross-domain neighborhood similarity distributions. DTS-Reg allows global reparameterization while preserving dynamics-relevant local topology, and integrates into Dreamer-style world models with minimal change. We formalize the method, freeze implementation hyperparameters, and specify a complete evaluation protocol on DMControl under viewpoint, appearance, and corruption shifts against DreamerV3 and augmentation / pointwise-alignment baselines. No measured ID/OOD performance numbers are claimed in this preprint; tabulated returns belong to a subsequent experimental campaign. A runnable PyTorch reference implementation is included as Appendix B.