In an emergency a human might ask a robot to fetch a wall‑mounted defibrillator down the hall. If the robot has mislabeled the device as a fire extinguisher, it will fail to locate the life‑saving equipment. MIT researchers introduced a new memory system called UQ‑DAAAM (Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment) that detects and corrects such labeling errors.\n\nUQ‑DAAAM compares multiple captions of the same object taken from different viewpoints, quantifies the vision‑language model’s (VLM) uncertainty, and, when the uncertainty exceeds a threshold, selects additional views to reduce it. An LLM then fuses the selected captions into a more accurate description, which is written back to memory.\n\nThe pipeline works as follows:\n1. Compute the VLM’s confidence for each generated word;\n2. Measure consistency by comparing this caption with those from other viewpoints, producing a single uncertainty score;\n3. If the score is too high, apply a heuristic that chooses the frames most likely to lower uncertainty—mathematically proven to identify optimal frames without re‑running the VLM, thus saving computation;\n4. Use an LLM to merge the captions, weighting newer and less ambiguous ones more heavily.\n\nTests showed that UQ‑DAAAM resolves ambiguities for about 92% of objects with only two extra captions, improving object‑description accuracy by 5% over the state of the art and yielding more reliable answers about object identity, location, and observation time.\n\nThe team plans to extend the approach to mobile manipulators and to let robots revisit high‑uncertainty areas or apply uncertainty quantification to future‑state prediction. The work is funded by the U.S. Army Research Laboratory and the Office of Naval Research.\n\nReview: By giving robots the ability to recognize and repair their own memory mistakes, UQ‑DAAAM brings a crucial “knowing‑what‑you‑don’t‑know” capability that can make autonomous systems safer and more trustworthy in real‑world tasks.