Modern autonomous driving systems (ADS) require real-time input from multiple sensors to make time-sensitive decisions using multiple deep neural network systems. This makes the correctness of those decisions crucial to the adoption of ADS as errors can cause significant loss. Sensors such as LiDAR for obstacle detections are sensitive to environmental changes and built-in inaccuracies and may fluctuate over time (i.e., frame to frame). While there has been extensive work to test ADS, it remains unclear whether the current ADS are robust against very subtle changes in LiDAR output data (i.e., point cloud data). In this work, we study the impact of the built-in inaccuracies in LiDAR sensors on LiDAR-3D obstacle detection models (i.e., deep neural network models) to provide insight into how they can impact the obstacle detection module (i.e., robustness) and by extension the trajectory prediction module (i.e., how the robustness of obstacle detection would impact on autonomous driving).

We propose a framework SORBET, that applies subtle perturbations to LiDAR data, evaluates the robustness of LiDAR-3D obstacle detection, and further assesses the impacts on the trajectory prediction module and ADS.
We applied SORBET to evaluate the robustness of five classic LiDAR-3D obstacle detection models, including one from an industry-grade Level 4 ADS (Baidu’s Apollo). Furthermore, we studied how the changes in the obstacle detection results would negatively impact trajectory prediction and ADS in a cascading fashion. Our evaluation highlights the importance of testing the robustness of LiDAR-3D obstacle detection models against subtle perturbations. We find that even very subtle changes in point cloud data (i.e., removing two points) may introduce a non-trivial decrease in the detection performance. Furthermore, we find that such a negative impact will further propagate to impact other modules in ADS, which may endanger the safety of ADS.