MIT's UQ-DAAAM lets robots flag and fix uncertain object descriptions in memory

MIT researchers led by Luca Carlone developed UQ-DAAAM, which scores how reliable a vision-language model's description of an object is by checking its confidence and consistency across viewpoints, then picks extra views when needed. It resolved ambiguity for about 92% of objects with two more captions each and improved description accuracy by 5%. The work will be presented at NeurIPS.

Summary written by DroidBazaar. The original article belongs to its publisher.