Remember with Confidence: Uncertainty Quantification for Spatio-temporal Memory with Probabilistic Guarantees
Researchers have developed a new method for quantifying uncertainty in spatio-temporal memory systems used by robots. This approach, called UQ-DAAAM, assigns an object-level semantic uncertainty score to VLM-generated captions, identifying unreliable descriptions. The system then actively refines these uncertain objects by selecting high-quality views and fusing captions to improve memory reliability and question-answering performance, as demonstrated on the OC-NaVQA benchmark. AI
IMPACT Enhances reliability of embodied AI systems by improving memory recall and reducing errors in object identification.