On November 5, 2026, the Massachusetts Institute of Technology (MIT) introduced UQ-DAAAM, a new method that lets robots correct their memories of previously observed objects using images taken from other viewpoints. The technology evaluates the uncertainty of descriptions that are stored on a 3D map and linked to each object's position and observation time, and updates them when necessary. By focusing additional image analysis on uncertain information, it improved accuracy in answering questions about objects. In experiments, uncertainty was also resolved for 92.3% of the objects judged to be uncertain. However, that figure is not the share of misidentifications that were correctly fixed. It is the share of memories whose uncertainty fell below a set threshold.
A weakness in robot object memory: mistaking a chimney for a squirrel
A robot mistakes a chimney for a squirrel and records that wrong information on its map. MIT's announcement uses this kind of example to explain the problem with robots' object memory. A recognition error does not stay a one-off failure: the wrong description is saved as a memory and can affect later question answering. For a robot to understand its surroundings over a long period, it needs a mechanism for revisiting recorded information when necessary.
The research builds on DAAAM, a system MIT introduced on June 17, 2026. It uses a vision-language model (VLM) to describe objects in images in text, then stores those descriptions linked to positions on a 3D map and observation times. This allows a robot to answer questions about when and where a particular object was. However, if an image is blurry or the surroundings are dark, an incorrect description may be recorded from the start. Distant objects and partially occluded ones are also hard to recognize. UQ-DAAAM adds a function that finds low-reliability information in memory and corrects it using additional observations.
The study's co-first authors are Harry Zhang and Nicolas Gorlo, and the senior author is Luca Carlone. According to the arXiv listing, the first version of the paper was posted on June 6, 2026, and MIT says it plans to present the work at NeurIPS. Details of the method and evaluation results are in the publicly available preprint and its appendix.
Measuring uncertainty from conflicting descriptions and choosing images to re-evaluate
UQ-DAAAM first selects three representative viewpoints for each object and compares the descriptions generated from each image. It converts each description into a semantic vector to measure how much the contents differ, and also calculates an inconsistency score based on word probabilities from when the description was generated. This combines the semantic spread across multiple descriptions with the uncertainty of each individual description. Unlike methods that simply ask a model whether it is confident, or that compare three descriptions by majority vote, it can quantitatively evaluate uncertainty for each object.
These results are used to select the objects that need additional image analysis. Rather than reprocessing every object uniformly, the aim is to avoid unnecessary processing by concentrating computing resources on highly uncertain memories. A high uncertainty score does not necessarily mean the object was actually misidentified, however. This stage does not confirm errors; it selects which memories should be re-evaluated.
Next, the system picks, from images that have already been captured, viewpoints from which the object can be seen more clearly. Re-analyzing every candidate image with a vision-language model would be computationally heavy, so the team used an existing viewpoint quality score. The score is higher when the target object appears larger and nearer the center of the image. Candidates are first narrowed down by quality score, and additional descriptions are generated for the selected images. This removes the need to re-analyze all candidate images, though it does not eliminate additional runs of the vision-language model altogether.
Importantly, the alternative-viewpoint images used in this experiment had already been acquired. The robot did not autonomously return to the scene to take new pictures; it selected suitable images from those already stored and re-evaluated its memory. Gathering new images through autonomous revisits is positioned as future work.
Finally, GPT-5-mini merges the existing description with the additional observations and updates the object's memory. In this step, the model is instructed to prioritize new observations while not adding information absent from the original description. Where early observations could not pin down a color and descriptions mix white and gray, multiple candidates are kept, such as "white or gray." Rather than forcing a single answer, the design records uncertain information as ambiguous.
Still, such instructions cannot completely prevent hallucination, in which an AI generates information that does not exist. Even so, separating the process that detects discrepancies among descriptions from the process that decides how to reflect them in memory differs from simply overwriting old descriptions with new ones.
Question-answering accuracy up 5.0 points; 92.3% of uncertainty resolved
In Table 1 of the paper, the research team presents a comparison of question-answering performance on the OC-NaVQA benchmark. All methods use the same GPT-5-mini as the question-answering model, so differences in accuracy can be separated from the effect of changing models. The questions evaluated concern descriptions of remembered objects. Results for the "All" category, which covers every evaluated item, are shown below.
| Method | Accuracy on description questions (All) | Difference from DAAAM |
|---|---|---|
| DAAAM | 71.1% | Baseline |
| DAAAM-RR (random additional processing) | 71.6% | +0.5 points |
| DAAAM-QR (additional processing based on viewpoint quality) | 73.8% | +2.7 points |
| DAAAM-ES | 74.3% | +3.2 points |
| UQ-DAAAM | 76.1% | +5.0 points |
Source: Table 1 of the paper. Compares the All category of OC-NaVQA, with GPT-5-mini used as the question-answering model throughout. Differences are percentage points, calculated by subtracting DAAAM's 71.1% from each method's accuracy.
UQ-DAAAM's accuracy was 76.1%, up 5.0 points from the original DAAAM's 71.1%. It also beat DAAAM-RR, which applies additional processing at random, at 71.6%, and DAAAM-QR, which selects images by viewpoint quality, at 73.8%. Against DAAAM-ES, the best-performing comparison method at 74.3%, however, the margin is only 1.8 points. Describing the result simply as a "5% improvement" would therefore leave unclear whether it is a relative improvement or a percentage-point difference, and which method it is compared against. The improvement over the original DAAAM and the advantage over other additional-processing approaches should be assessed separately.
Another key metric is the uncertainty resolution rate (URR). It represents the share of objects initially judged uncertain whose uncertainty score fell below a set threshold after additional processing. With a cap of two additional queries, UQ-DAAAM reached a URR of 92.3%. By comparison, the viewpoint-quality-based selection method reached 34.2% and the random selection method 5.6%. Note that this was not a condition where two extra images were added uniformly for every object; the objects to process and the number of additional queries were determined by uncertainty.
With a cap of one additional query, the URR was 68.9%, and with three it was 94.4%, so allowing more queries raised the share of objects whose uncertainty could be resolved. However, the 92.3% figure does not mean that 92.3% of misidentifications were correctly fixed. It is only the share whose uncertainty score fell to or below the threshold, and reduced disagreement among descriptions does not guarantee the object is recognized correctly.
Images processed at 13.5 Hz, but that is not the same as time to finish correcting memory
For the implementation performance evaluation, the team used seven sequences from the CODa dataset, totaling 98,866 frames. The setup was a single NVIDIA RTX 4090 with 24 GB of memory, and it recorded a processing speed of 13.5 Hz against a 10 Hz input. DAM-3B was used to generate object descriptions, and GPT-5-mini was used for merging descriptions and question answering. The result indicates that, at least in the evaluation environment, the main processing pipeline ran faster than the rate at which input frames arrived.
However, the 13.5 Hz figure cannot be interpreted as meaning a robot can complete all memory corrections instantly while moving. MIT's announcement describes real-time updates during movement, but according to the paper's appendix, the additional vision-language model queries are run asynchronously at shutdown. In other words, the speed at which input images are processed differs from the time it takes to finish the extra queries and complete the memory correction. The implementation also sets a 300-second timeout, but that is not a measured value of the time corrections took. Evaluating real-time capability requires checking not just image-processing throughput but also the delay until memory updates are complete.
Specific results are also reported on how objects were selected for processing. Of the 18,298 objects evaluated, 6,324, or 34.6% of the total, were chosen for additional processing. For those objects, the number of additional viewpoints used averaged 1.32 ± 0.47. The design of concentrating computing resources on highly uncertain memories, rather than applying the same number of additional passes to every object, is reflected in the actual allocation of processing.
Meanwhile, accuracy on position and time questions did not improve in every evaluation category. In the All category, position error fell from the original DAAAM's 41.75 m to 37.84 m, and time error from 1.792 minutes to 1.589 minutes. In the Long category, however, position error worsened from 53.47 m to 58.94 m. So even when the overall average improves, performance can decline under particular conditions. These are errors in answers about the positions and observation times of remembered objects, not measurements of the robot's own localization accuracy or collision-avoidance performance.
Even when descriptions from multiple images agree, misidentifications can remain
The study also found cases where object misidentification persisted even after uncertainty dropped. The paper's appendix shows an example in which a bike-parking rack continued to be recognized as a phone booth even after clear images were added. Additional processing reduced disagreement among descriptions and lowered the uncertainty score, but the stored description remained wrong. This illustrates an important limitation of the method: reducing a memory's uncertainty does not necessarily make its content correct.
Even when descriptions from multiple viewpoints agree, there is no guarantee they are correct. If the underlying vision-language model repeats the same kind of misidentification, adding images may leave the wrong description in place. Questions also remain about how well the uncertainty score reflects actual recognition accuracy, that is, about confidence calibration. In addition, extra model queries and description merging carry computational costs, and with long-term operation the amount of memory stored, including past observation history, will grow.
The paper also offers theoretical guarantees for viewpoint selection, but they depend on assumptions such as stochastic dominance in the ranking of viewpoint quality and expected information gain. Selecting images with high viewpoint-quality scores therefore does not guarantee that a misidentification will be corrected. The evaluation is also based on datasets the research team used; it does not verify success rates in real-world object manipulation or robot safety in homes.
The team aims next to add a function for robots to autonomously return to highly uncertain places and re-observe them, and to apply the method to robots that combine movement with object manipulation. At that stage, what will matter is not only whether new observations can correct misidentifications, but also whether memory updates finish in time for the robot to act. It will also need to be verified whether corrected memories actually improve manipulation success rates. UQ-DAAAM shows an effective way for robots to revisit past recognition results, but whether that can be translated into real-world behavior will be the focus going forward.