ROSBENCH: A Simulation-Based Benchmark for Sensor Quality and Environmental Conditions Robustness in AV Perception

The perception capabilities of autonomous vehicles (AVs) rely on high-quality sensor data to accurately interpret the environment. Among the key sensing modalities, LiDAR and RGB cameras offer distinct advantages. LiDAR provides precise depth estimation and object localization, while cameras capture rich visual details. However, their effectiveness depends on factors such as resolution, measurement accuracy, and environmental conditions, making a systematic comparison essential for optimizing AV perception. Sensor fusion, which integrates multiple sensing modalities, can improve robustness by mitigating the limitations of individual sensors. The growing reliance on simulation-based research has accelerated AV development, with platforms like CARLA providing scalable, cost-effective environments to evaluate sensor performance under controlled yet diverse conditions, including adverse weather and complex traffic scenarios. In this work, we propose a comprehensive and robust simulated benchmark ROSBENCH for evaluating the robustness of Perception and Prediction systems. This benchmark can be used as a consistent reference point for validating the impact of environmental and sensor conditions on vision algorithms and is also easily configurable and thus easily adaptable.