Learning what dynamics changes matter, not what they are

Robot policies typically break when physical conditions change—friction shifts, payload adds weight—because they're trained for one specific environment. Rather than explicitly tagging what changed (the traditional approach), this work lets policies learn which dynamics shifts actually matter for task success, using contrastive learning to build a clean latent space. On MuJoCo tasks, the method outperforms parameter-centric baselines under severe and compound changes while maintaining in-distribution performance.