Primary domain: Robotics & Automation (Autonomous Systems)
Secondary domain: Artificial Intelligence & Machine Learning (Computer Vision)
Article type: Research & Discovery
Technology: UQ-DAAAM uncertainty-aware robot memory
Innovation maturity: Experimental
Mission: Research
Business impact: Could make household and industrial robots more reliable at multistep tasks by flagging and fixing uncertain perceptions
Audience: Research, IT, Business/Leadership
Publication date: November 2026
Author / Contributor: Innovation team
Related technologies: Vision-language models; large language models; DAAAM; uncertainty quantification; MIT SPARK Lab; physical AI
Picture a robot sent to fetch a wall-mounted defibrillator. If its stored memory wrongly lists that object as a fire extinguisher, it may never find the device. That example opens an MIT News report, published November 5, 2026, on a new system that lets a robot notice when its own memory might be wrong and repair it.
The system is called UQ-DAAAM, short for Uncertainty Quantification for Describe Anything, Anywhere, at Any Moment. It was developed by researchers at MIT, led by associate professor Luca Carlone, and will be presented at the Conference on Neural Information Processing Systems.
The problem: confident but wrong captions
UQ-DAAAM builds on the team’s earlier DAAAM project, a robot memory that uses a vision-language model (VLM) to describe objects in scenes captured by onboard cameras. Those descriptions are saved as detailed captions in a 3D, map-based representation of the surroundings, which the robot can query later to answer questions.
The weak point is perception. A robot may see an object only partially, from far away, or in a blurry or dimly lit frame. In those cases the VLM can describe it incorrectly, for example by mistaking a gray chimney for a squirrel on a roof. Without a way to correct itself, as co-lead author Harry Zhang put it, the wrong memory would persist as the robot continues through its environment and could affect later tasks.
How the correction works
While moving, the robot constantly collects images, and many show the same object from different angles. UQ-DAAAM uses those extra viewpoints in two steps, according to MIT.
First, it measures how confident the VLM is in the words it generated for a suspicious object. Second, it checks consistency by comparing that description with captions of the same object from other vantage points. The two sources of doubt are fused into one number that reflects how certain the model is about what it saw.
If that number passes a threshold, the system refines the caption. A simple rule, or heuristic, picks which additional views will most reduce uncertainty; it might choose a frame a few seconds later that clearly shows smoke coming from the chimney. The researchers say they proved mathematically that the heuristic can pick the best frames without rerunning the VLM, which sharply cuts computation. A large language model then merges the chosen captions into a single description, giving more weight to newer and less ambiguous ones.
What the tests showed
MIT reports that this refinement resolved ambiguities for about 92 percent of the objects in the robot’s environment while needing only two additional captions per object. It improved the accuracy of object descriptions by 5 percent compared with the state of the art. The system also answered questions about the objects a robot observed, where it saw them and when it last encountered them more accurately than baseline methods.
The speed requirement matters as much as accuracy. Carlone noted that the challenge for co-lead authors Harry Zhang and Nicolas Gorlo was combining VLMs with uncertainty quantification so it runs in a fraction of a second on a robot that is moving. Gorlo said doing this across the whole map improves the robot’s understanding of what it has observed.
What it means for physical AI
Carlone drew a pointed contrast with chatbots. For physical AI, he said, it is not enough to attach a disclaimer that the robot might be wrong; allowing a robot to know what it does not know will be essential for deploying the technology safely. That is the real contribution here: not a bigger model, but a way for a machine to attach honest confidence to what it remembers.
If the approach carries over, it could make household or industrial robots better at complex, multistep jobs such as clearing a dining room table or helping assemble a piece of machinery alongside people. The work is still research, and MIT does not describe a product or deployment timeline.
Next, the team wants to apply UQ-DAAAM to other kinds of robots, including mobile manipulators, and to other domains. They are also studying how to flag high-uncertainty areas a robot could revisit to improve its understanding, and exploring uncertainty quantification in systems that predict a robot’s future surroundings.
The bottom line
UQ-DAAAM tackles a quiet failure mode of AI-powered robots: a memory that is wrong and never gets fixed. By letting a robot weigh its own doubt and gather a few extra views, MIT’s team shows a practical route to more trustworthy machine memory. It is an early research result, but it points at a design principle that builders of physical AI are likely to adopt.