Researchers have developed an algorithm to improve the safety of autonomous systems with imperfect perception. These systems can make unsafe decisions when sensor readings are misclassified. The new approach uses confidence intervals for perception outcomes and models the system as a finite Interval Partially Observable Markov Decision Process. This enables a runtime shield that guarantees safe actions within a certain time horizon, with high probability over training data. Experiments on four case studies show improved safety compared to state-of-the-art baselines. --- Why it matters: This matters because autonomous systems with imperfect perception are increasingly common in applications like self-driving cars and robots. Engineers can use this algorithm to improve the safety of these systems, reducing the risk of accidents and damage. Source: https://arxiv.org/abs/2604.20728