Enhancing Robustness Testing for DNN-Controlled Robotics Systems
A new approach improves the transferability of robustness tests from simulation to physical robots.
Deep Neural Networks (DNNs) are increasingly used in Cyber-Physical Systems (CPSs), but small input changes can lead to unsafe behaviors. This study introduces a multi-objective evolutionary method that generates effective robustness tests across various operational scenarios.
Methodology
The proposed method utilizes explainability-guided techniques to create sparse perturbations based on visual and behavioral clustering of images. The effectiveness of this approach was tested on a DNN-controlled LeoRover in a simulated environment and validated on a physical robot.
Results
The results showed a median success rate of 70.0% for the new method, outperforming unguided searches. The study also found a strong correlation between simulated and physical failure times, highlighting the importance of physical validation.