MIT researchers have built a memory system for robots that flags entries that may be wrong and corrects descriptions of mislabelled objects, according to the university. The system is called UQ-DAAAM, short for Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment.
It compares several descriptions of the same object seen from different angles to measure how uncertain the robot is about what it saw, picks more views to reduce that uncertainty, and combines the captions into a more accurate description saved in memory. The researchers made the process fast enough for a mobile robot to update uncertain entries as it moves.
Compared with baseline methods, UQ-DAAAM made the robot more confident and more accurate when answering questions about objects, where they were and when it saw them, the researchers report. The paper's senior author, MIT associate professor Luca Carlone, said robots need to know what they don't know to be deployed safely. The study is on the arXiv preprint server.
