Reproduce “PEM: Perception Error Model for Virtual Testing of Autonomous Vehicles” on OpenResearch

About this paper

Andrea Piazzoni, Jim Cherian, Justin Dauwels, Lap-Pui Chau

Researchers developed Perception Error Models (PEMs) that approximate the sensing and perception system of autonomous vehicles by directly modeling output errors. These models enable efficient, safety-centric virtual testing, showing how different sensor configurations lead to varying collision rates in urban driving scenarios.

Key insights

• Directly modeling perception errors at the output level provides a computationally efficient and robust method for evaluating AV safety in virtual environments, overcoming limitations of both physics-based simulations and traditional perception metrics.
• The safety criticality of a perception error is highly context-dependent, varying with the driving scenario and its effect on the AV's driving policy, which is not captured by conventional metrics.
• Incorporating contextual factors like object distance, angle, and occlusion into PEMs is crucial for accurately representing real-world perception performance variations and enabling meaningful safety assessments.

Summary by alphaXiv.