MIT's UQ-DAAAM fixes robot memory errors

MIT researchers have developed UQ-DAAAM, a system that allows robots to identify and correct errors in their visual memory. The method quantifies uncertainty in object descriptions generated by vision-language models, prompting the robot to seek additional viewpoints when confidence is low. This process improves description accuracy by 5% over state-of-the-art baselines and resolves ambiguities for 92% of objects using only two extra captions. The research, presented at NeurIPS, aims to make autonomous agents safer and more reliable in dynamic environments.