MIT proposes UQ-DAAAM to correct erroneous object descriptions in robot memory using multi-view images

MIT researchers propose the UQ-DAAAM system, which quantifies a robot's uncertainty about what it sees by comparing multiple descriptions of the same object from different viewpoints, selects additional viewpoints to correct erroneous annotations, and writes the corrections back to memory. The method adds uncertainty quantification to the DAAAM framework, uses heuristic rules to select frames that can reduce uncertainty without rerunning the vision-language model, and then uses a large language model to fuse the multiple descriptions.