MIT Researchers Build UQ-DAAAM to Fix Robot Memory Errors, Lifting Object Accuracy 5%
Updated
Updated · MIT News · Nov 5
MIT Researchers Build UQ-DAAAM to Fix Robot Memory Errors, Lifting Object Accuracy 5%
3 articles · Updated · MIT News · Nov 5
Summary
UQ-DAAAM lets robots flag uncertain memory entries and rewrite mislabeled object descriptions in real time, addressing failures such as confusing a defibrillator with a fire extinguisher.
The MIT system measures uncertainty by comparing captions of the same object from multiple viewpoints, then selects the most useful extra frames without rerunning the vision-language model.
Using an LLM to fuse those captions, the method resolved ambiguities for about 92% of objects while needing only two additional captions per object.
Tests showed a 5% gain in object-description accuracy and more accurate answers about what the robot saw, where it saw it and when it last encountered it.
The work, to be presented at NeurIPS, is aimed at safer household and industrial robots that can handle multistep tasks and collaborate more reliably with humans.