Remember with Confidence: Uncertainty Quantification for Spatio-temporal Memory with Probabilistic Guarantees

1 min read

This paper adds object-level semantic uncertainty to VLM-based spatio-temporal memory, scored as how much an object’s captions disagree across views. UQ-DAAAM uses the score to refine uncertain objects within a fixed query budget, choosing good views and fusing their captions into one description, and the authors prove that better candidate views are more likely to reduce uncertainty. On OC-NaVQA it reduces uncertainty more and answers spatio-temporal questions better than the baselines.

Relevance

A robot memory that stores VLM captions as facts will keep wrong facts for a long time. This builds on DAAAM and pairs with Worth Remembering for long-term autonomy and 3D scene graphs.