Drive in Corridors: Enhancing the Safety of End-to-end Autonomous Driving via Corridor Learning and Planning

TL;DR

End-to-end autonomous driving struggles with safety due to lacking explicit behavioral constraints. The authors propose corridor learning, introducing a geometric corridor as an intermediate representation that is learned and enforced within a differentiable trajectory optimization framework. They define a BEV-based corridor as a sequence of rectangles, annotate corridors via a maximum empty rectangle approach from obstacles, and train a multi-task network with corridor-specific losses, followed by a differentiable QP planner that respects corridor constraints. Across nuScenes and Bench2Drive, the method achieves substantial reductions in agent and curb collisions and improves closed-loop success, demonstrating enhanced safety and interpretability while highlighting challenges in gradient stability for optimization. This work advances practical, interpretable end-to-end driving by tightly coupling learned perception and corridor-aware planning through differentiable optimization.

Abstract

Safety remains one of the most critical challenges in autonomous driving systems. In recent years, the end-to-end driving has shown great promise in advancing vehicle autonomy in a scalable manner. However, existing approaches often face safety risks due to the lack of explicit behavior constraints. To address this issue, we uncover a new paradigm by introducing the corridor as the intermediate representation. Widely adopted in robotics planning, the corridors represents spatio-temporal obstacle-free zones for the vehicle to traverse. To ensure accurate corridor prediction in diverse traffic scenarios, we develop a comprehensive learning pipeline including data annotation, architecture refinement and loss formulation. The predicted corridor is further integrated as the constraint in a trajectory optimization process. By extending the differentiability of the optimization, we enable the optimized trajectory to be seamlessly trained within the end-to-end learning framework, improving both safety and interpretability. Experimental results on the nuScenes dataset demonstrate state-of-the-art performance of our approach, showing a 66.7% reduction in collisions with agents and a 46.5% reduction with curbs, significantly enhancing the safety of end-to-end driving. Additionally, incorporating the corridor contributes to higher success rates in closed-loop evaluations. Project page: https://zhiwei-pg.github.io/Drive-in-Corridors.

Figures (4)

• Figure 1: Architecture of our method. The multi-task backbone processes multi-view images as input and outputs perception results, the reference trajectory, and the corridor. The predicted corridor is supervised using the annotation from the dataset and further refined through the safety loss to minimize overlap with agents and curbs. Built on the kinematic bicycle model, the differentiable optimization module utilizes the corridor as the constraint and is aimed to track the reference trajectory. Finally, the optimized trajectory is trained to imitate human driving actions.
• Figure 2: Corridor Annotation. (a) Agent boxes and map data are loaded from the dataset. (b) Coordinates are sampled. (c) The largest rectangle is identified and highlighted on top and other candidate rectangles are displayed below ground level with high opacity. (d) Iterating over timestamps and combining the results forms the corridor, where rectangles are stacked on the BEV, and colors represent different timestamps.
• Figure 3: Visualization comparison of the learned corridors with auxiliary losses. Left: Corridors learned using only $\mathcal{L}_{cor}$. Right: Auxiliary losses are incorporated, which reduces overlap with curbs. Ground-truth maps and agents are displayed for clearer comparison. Minor intersection between the predicted corridor and ground-truth map may occur due to imperfections in perception outputs, also seen in Fig \ref{['fig:qualitative']}. Note that the bottom case follow the left-hand traffic rules.
• Figure 4: Qualitative Results. Each subfigure presents the perspective view with projected trajectory and corridor, and the 'PREDICTION' results (perception outputs with the reference trajectory) alongside the 'GROUND TRUTH' (ground-truth maps and agents and the optimized trajectory). Subfigures (c) and (d) highlight cases where the reference trajectory collides, but the corridor constraints successfully guide the optimized trajectory to remain safe. Slight discrepancies between the perspective and BEV view may occur due to the estimated corridor height.