Robustness Verification of a Reinforcement Learning-Based Agent for Automated Car Parking

Safety critical systems require careful verification of their performance in well-defined operational conditions. We have verified the performance of an automated driving (AD) parking agent trained through a reinforcement learning-based automated driving (AD) parking agent in three exemplary critical scenarios in the CARLA driving simulation environment. Results, obtained through the Marabou formal verification tool, show that the system does not produce outputs in a range that would violate expert-defined safety assumptions. Analyzing robustness to input signal perturbation, we observe that injection of Gaussian and pixel noise decreases slightly the safety performance. Interestingly, especially at higher noise levels, the agent frequently decides to remain stationary, which would allow the system to safely issue a take-over request to the driver. These results indicate significant further research in the field.