NPPC: Neural Parametric Planning Cost for End-to-End Autonomous Driving

Abstract

Recent advances in end-to-end autonomous driving (E2E AD) often rely on direct expert trajectory imitation to train planners. While promising, these methods are susceptible to biases and compounding errors, which can compromise safety and road compliance. Inverse Optimal Control (IOC) has the potential to guide the planner and mitigate trajectory biases by inferring the underlying expert behavior as a planning cost. However, existing IOC approaches in E2E AD face significant limitations: expert behavior is modeled using manually designed cost terms or spatially discrete cost-maps, which may not accurately capture expert reasoning, or through high-dimensional neural networks to score trajectory proposals, restricting planners to candidate-based approaches. To address these limitations, we propose Neural Parametric Planning Cost (NPPC), an IOCbased approach that represents expert behavior through a lowdimensional, spatially continuous parametric cost function. NPPC effectively captures expert-like evaluation of the current driving scene without manual engineering or discretizations. Moreover, NPPC's low-dimensional representation enables direct optimization for planning, allowing maneuver generation beyond the predefined trajectory candidates across diverse driving scenarios. NPPC achieves superior performance over the previous state of the art in vision-based E2E AD, reducing the average collision and curb collision rates by 5.7% and 9.8%, respectively, in the open-loop evaluation and improving the driving score by 3.5% in the closed-loop evaluation.