MIT researchers build system that lets robots flag and fix unreliable object labels in memory

Researchers at the Massachusetts Institute of Technology have developed a memory system that identifies likely errors in a robot's stored object descriptions and corrects them, MIT News reported on Nov. 5.

The system is called UQ-DAAAM, short for Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment. The team said it could eventually help household or industrial robots handle complex, multistep jobs such as clearing a dining room table or helping people assemble machinery.

The problem: wrong memories that persist

UQ-DAAAM builds on the group's earlier system, DAAAM. That system uses a vision-language model, an AI model that links images and text, to describe objects seen by a robot's onboard cameras. The descriptions are saved as captions in a 3D map of the environment that the robot can later query.

The researchers said a robot may see an object only partly, from far away, or in a blurry or dim image, and DAAAM can then describe it wrongly. One example given was a gray chimney mistaken for a squirrel on a rooftop.

“With no way to correct itself, this wrong memory would persist as the robot proceeds in its environment, which could impact downstream applications,”

Harry Zhang, MIT graduate student and co-lead author

How the correction works

As a robot moves, it gathers many images of the same objects from different angles. According to MIT, UQ-DAAAM first calculates the model's confidence in the words it produced for an object, then checks how consistent that description is with captions of the same object from other vantage points.

“We fuse both sources of uncertainty into a single number that captures how certain the VLM is in its overall understanding of the object,”

Luca Carlone, associate professor in MIT's Department of Aeronautics and Astronautics

If that number passes a threshold, the system picks extra views using a simple rule, or heuristic. The researchers said they proved mathematically that the rule can identify the best frames without rerunning the vision-language model, sharply reducing computation. A large language model then merges the selected captions into one description, giving more weight to newer and less ambiguous captions.

Reported results

MIT said the refinement step resolves ambiguities for about 92 percent of objects in the robot's environment while needing only two additional captions per object. It raised the accuracy of object descriptions by 5 percent compared with the state of the art, according to the report.

Compared with baseline methods, the system also gave more accurate answers to questions about which objects the robot observed, where it saw them and when it last encountered them, MIT said.

“Contrary to ChatGPT, when we think about physical AI, it is not enough to simply put a disclaimer cautioning that sometimes the robot might be wrong. Allowing a robot to know what it doesn’t know is going to be essential for this technology to be deployed safely and have a positive impact on society,”

Luca Carlone, associate professor, MIT, senior author of the paper

Next steps

Carlone is the senior author of the paper, with Zhang and Nicolas Gorlo as co-lead authors. The work will be presented at the Conference on Neural Information Processing Systems.

The team plans to use UQ-DAAAM to guide other robot types, such as mobile manipulators, and to apply its uncertainty techniques in other areas. It is also studying how the system could point robots to high-uncertainty areas to revisit, and looking at uncertainty methods for systems that forecast how a robot's surroundings will change. Funding came partly from the U.S. Army Research Laboratory and the U.S. Office of Naval Research, MIT said.