AI Robot Mistakes a Chimney for a Squirrel in MIT Experiment

An artificial intelligence system mistook a chimney for a squirrel during a robotics experiment, demonstrating how easily machines can misunderstand the world around them. Researchers at the Massachusetts Institute of Technology have now developed a way for robots to recognise uncertain information and correct mistakes stored in their memory.

The research, announced on 5 November 2026, could help future robots perform tasks more reliably in homes, workplaces and other environments. It also addresses a fundamental problem with artificial intelligence, which can produce convincing descriptions even when those descriptions are incorrect.

How did a chimney become a squirrel?

Robots equipped with cameras can photograph their surroundings and use artificial intelligence to identify objects. Those descriptions may then be stored in a digital map, allowing the robot to remember what it has encountered and where particular objects are located.

However, a camera might capture something from an unusual angle, in poor lighting or at a considerable distance. In MIT's demonstration, the system incorrectly identified a dark grey chimney on a rooftop as a squirrel.

The mistake might appear amusing, but incorrect memories can cause practical problems. A robot asked to locate a particular object could struggle if the item had previously been identified incorrectly.

Teaching a robot to question its memory

Researchers Harry Zhang and Nicolas Gorlo, working with senior author Luca Carlone, developed a system called UQ-DAAAM. It assesses uncertainty in the descriptions stored by a robot and identifies information that may need to be corrected.

The system considers how confident its image-recognition model is and whether descriptions of the same object agree when viewed from different angles. When uncertainty is sufficiently high, the robot uses additional observations to produce a more reliable description.

In the chimney experiment, the system was able to use further views to recognise the object correctly. Rather than preserving the original mistake, it updated the information stored in memory.

The researchers reported that their approach resolved ambiguities for approximately 92% of objects in the experimental environment, using only two additional captions per object. It also improved object-description accuracy by 5% compared with the previous state-of-the-art approach.

That 92% figure describes the success of the ambiguity-resolution process, not the robot's overall recognition accuracy. The results relate to the experimental setting rather than a guarantee of performance in every real-world environment.

Could this make future robots safer?

Recognising uncertainty is particularly important when artificial intelligence controls physical machines. An incorrect answer from a chatbot may be inconvenient, but a robot acting on mistaken information could attempt to collect the wrong object or move towards an unsuitable location.

MIT researchers suggest that the technology could eventually support robots carrying out more complex tasks, including clearing tables or assisting with machinery. Those applications remain potential future uses rather than capabilities demonstrated by this particular experiment.

The team is also investigating whether robots could identify areas they need to revisit to improve their understanding of their surroundings. That would allow machines to gather more information instead of relying entirely on their first observations.

The chimney and squirrel mistake provides a memorable example of a serious technical challenge. Making robots more dependable may require teaching them not just how to recognise objects, but also when to question what they think they know.

Research and sources

Harry Zhang, Nicolas Gorlo and Luca Carlone, Massachusetts Institute of Technology, 5 November 2026. MIT research findings and experimental imagery.