UF Research Finds Simple Visual Patterns Can Mislead AI-Powered Vehicles and Robots

Summary

University of Florida researchers found that simple repeated patterns, such as stripes or checkerboard-like designs, can cause autonomous vehicles and robots to misjudge how far away an obstacle is. The problem affects systems that estimate depth from stereo cameras, whether they rely on traditional algorithms or artificial intelligence models. Tests showed that the same underlying weakness appeared across multiple sensors, algorithms and AI models, suggesting it is not confined to one manufacturer or application. In a driving demonstration, researchers projected a pattern onto the back of a vehicle, causing the perception system to interpret part of the vehicle as closer than it was and trigger a response such as braking. A small, strategically placed pattern was sufficient; an attacker would not need to cover the entire scene. The researchers said the patterns can also occur naturally, including on fences or regularly spaced objects, making the issue a safety concern even without a deliberate attack. Similar depth errors could cause sudden maneuvers or collisions for drones and ground robots navigating around buildings, vehicles, birds and other obstacles. The team developed defenses aimed at the underlying sensing problem: software for traditional depth-estimation systems to detect repeated-pattern conditions, and model modifications for AI-based systems. They evaluated the approach in simulations and with a real vehicle and sensors at the University of Florida’s controlled testing facility. The study is scheduled for presentation at the ACM Conference on Computer and Communications Security in November.