Researchers at MIT have developed a system that lets robots detect when their memory of an object is unreliable and correct it. MIT News described the work on November 5, 2026. The system, called UQ-DAAAM (Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment), extends the team's earlier DAAAM robot memory system.
UQ-DAAAM estimates how uncertain a robot is about an object's description by combining the vision-language model's own confidence with how consistently the object is described from different viewpoints. When uncertainty is high, it selects additional views that should reduce it and uses a language model to merge the resulting captions into a more accurate description.
To keep this fast enough for a moving robot, the researchers use a heuristic to choose the extra views without rerunning the vision-language model, a shortcut they prove mathematically. According to MIT, the refinement step resolves ambiguities for about 92% of objects using only two additional captions per object, and improves object description accuracy by 5% over the previous state of the art. Answers about what objects are, where they are and when they were last seen were also more accurate than baseline methods.
The work was led by Luca Carlone, an associate professor in MIT's Department of Aeronautics and Astronautics, with co-lead authors Harry Zhang and Nicolas Gorlo. It is presented at NeurIPS in the paper "Remember with Confidence: Uncertainty Quantification for Spatio-temporal Memory with Probabilistic Guarantees" (arXiv:2606.08277), and was supported in part by the US Army Research Laboratory and the Office of Naval Research.
For teams building robots or agents on top of vision-language models, the practical idea is to treat captions as estimates with confidence attached rather than ground truth, and to spend extra perception only where that confidence is low.