Deep neural networks are increasingly used to control robotic navigation systems, but small perturbations to their inputs can lead to unsafe system-level behavior. According to a new arXiv preprint, existing approaches to robustness testing often focus on optimizing perturbations for individual scenarios, which may not transfer well across different conditions.
The paper, titled "Generalizable Robustness Testing of DNN-Based Robotic Navigation Systems via XAI-Guided Search," proposes using explainable AI (XAI) to guide the search for effective perturbations. This approach aims to identify weaknesses in a more generalizable way, rather than tailoring tests to a single case.
The abstract is brief, so details on the method and experimental results are not yet available. The work highlights a growing interest in combining explainability with robustness testing for safety-critical robotic systems.