Clearing up a robot’s fuzzy memory
Enabling physical AI systems to recognize and correct their own visual perception errors in real time is critical for safely deploying autonomous robots in complex human environments.
Reported by MIT News ·
What happened
MIT researchers have developed UQ-DAAAM, a real-time memory system for robots that uses uncertainty quantification to identify and correct mislabeled object descriptions generated by vision-language models. By comparing object captions from multiple viewpoints and fusing them with a large language model, the system resolves ambiguities for about 92 percent of objects and improves overall description accuracy by 5 percent.
Who it’s about
- MIT
- UQ-DAAAM
- Luca Carlone
- Harry Zhang
- Nicolas Gorlo
Source
MIT News
Clearing up a robot’s fuzzy memory
More on robotics
Get a brief picked for you
Sign in to follow topics and companies, save stories and search the archive.
Sign in