Worth Remembering uses Bayesian surprise to decide which episodes a robot keeps in long-term memory, computed in the latent space of V-JEPA 2 so it does not depend on the deployment. The surprise-gated episodic memory augments a 4D scene-graph spatial memory. It beats prior robot memory methods by at least 12% on temporal, spatial and binary robot question answering, and an unsupervised causal version outperforms supervised and non-causal methods on event segmentation.
Relevance
Surprise gives a deployment-agnostic rule for what to keep when future tasks are unknown. It builds on V-JEPA 2 and pairs with UQ-DAAAM for long-term autonomy; see also dynamic scenes.
Links
- arXiv
- Subfield: Semantic mapping, Perception