MIT system lets robots find and fix wrong object labels in memory

Why a robot's wrong memory is a problem

MIT offers an example: in an emergency, a person might ask a robotic assistant to fetch a wall-mounted defibrillator down the hall. If the robot has wrongly labeled the defibrillator as a fire extinguisher in its stored memory, it may not know where to find the device.

The new system builds on the team's earlier work, DAAAM. That robot memory system uses a vision-language model (VLM) to describe objects in scenes captured by the robot's onboard cameras. The descriptions are stored as detailed captions in a 3D map-based representation of the environment, which the robot can call upon later to answer questions.

The weakness is that a robot sometimes sees objects only partially or from far away, or its camera captures blurry or poorly lit images. In such cases DAAAM might describe an object incorrectly, for example by mistaking a gray chimney for a squirrel perched on a rooftop. Harry Zhang, co-lead author, says that "with no way to correct itself, this wrong memory would persist as the robot proceeds in its environment, which could impact downstream applications."

How UQ-DAAAM works

UQ-DAAAM stands for Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment. It adds uncertainty quantification to the DAAAM framework, so the robot can measure its confidence in its object descriptions and fix incorrect captions as it travels.

As the robot moves, it collects images, many of which show the same object from different angles. The system first calculates the VLM's confidence in the words it generated for a potentially mislabeled object. It then measures consistency by comparing that description with captions of the same object from other vantage points. Luca Carlone, senior author of the paper, explains: "We fuse both sources of uncertainty into a single number that captures how certain the VLM is in its overall understanding of the object."

If that number exceeds a threshold, the uncertainty is considered too high and the system refines the caption using additional images. A heuristic, or simple rule, decides which extra viewpoints will reduce uncertainty. It might, for instance, pick an image from a few frames later that clearly shows smoke coming from the chimney. The researchers say they mathematically proved that this heuristic can identify the best frames without rerunning the VLM, which dramatically reduces computation.

Once the useful captions are chosen, an LLM fuses them into a single object description, giving more weight to new captions and those with less ambiguity.

Results and next steps

According to MIT, this refinement step resolves ambiguities for about 92 percent of the objects in the robot's environment while requiring only two additional captions per object. It increased the accuracy of object descriptions by 5 percent compared to the state of the art. The system also gave more accurate answers about observed objects, where they were seen and when they were last encountered than baseline methods.

Speed was a central challenge. Carlone says the difficulty was combining VLMs with uncertainty quantification "into something that is running in a fraction of a second on a robot that is moving."

The researchers next want to use UQ-DAAAM to guide other types of robots, such as mobile robotic manipulators, and apply the uncertainty techniques to other domains. They are also studying how the system could identify areas of high uncertainty that a robot could revisit, and exploring uncertainty quantification in systems that predict the future state of a robot's environment.

The work will be presented at the Conference on Neural Information Processing Systems. Zhang and Nicolas Gorlo, both MIT graduate students, are co-lead authors. The research is supported in part by the U.S. Army Research Laboratory and the U.S. Office of Naval Research.

What it could mean for users

MIT says the method could eventually make household or industrial robots better at complex, multistep tasks, such as clearing a dining room table or working with human collaborators to assemble a piece of machinery. Carlone argues that for physical AI, a disclaimer that the robot might be wrong is not enough. "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," he says.