PHYSICAL AI
MIT Teaches Robots to Admit “I’m Not Sure What I Saw”
November 5, 2026 · Updated November 8, 2026
#PhysicalAI #AIVision #Research
Imagine asking a robot to fetch the defibrillator down the hall, and it confidently brings you a fire extinguisher. Researchers at MIT want to make sure robots know when their memory might be wrong. Their new system is called UQ-DAAAM, short for Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment.
It builds on DAAAM, an earlier MIT system that uses a vision-language model to describe objects a robot sees and stores those descriptions in a 3D map of its surroundings. The problem: bad lighting, distance or a blurry view can produce wrong descriptions, and those mistakes stay in memory.
How it fixes the problem
- Measure doubt: it checks how confident the AI was in its description
- Compare views: it compares descriptions of the same object from different angles
- Look again: when uncertainty is high, it picks extra views that should clear things up
- Rewrite memory: it merges the best descriptions into one corrected entry
The refinement resolved ambiguities for about 92% of objects while needing only two extra captions per object, and improved description accuracy by about 5% over the previous best method. The work will be presented at the NeurIPS conference.
Why it matters for robotics
A robot that knows what it doesn’t know is a safer robot. As mobile robots and mobile manipulators take on fetch-and-carry tasks in plants and warehouses, reliable memory of where things are becomes just as important as a strong gripper.