MIT (Luca Carlone's group; co-lead authors Harry Zhang and Nicolas Gorlo) · Uncertain Robot Memory

UQ-DAAAM

UQ-DAAAM adds uncertainty quantification to DAAAM, a robot memory system that uses a vision-language model to caption objects in a 3D map. It measures the robot's confidence by combining the VLM's own token-level uncertainty with caption consistency across views, then selects extra viewpoints to refine unreliable captions. The authors report that it improves confidence and answer accuracy against baseline methods, and say a heuristic for choosing frames is mathematically proven to work without rerunning the VLM. The page gives no code or licence details, and the work will be presented at NeurIPS.