Real-World Evaluation of two Cooperative Intersection Management Approaches

TL;DR

The paper evaluates two cooperative intersection management approaches—an optimization-based, multi-scenario predictor and a graph-based RL policy—within mixed traffic at urban unsignalized intersections. It extends a realistic simulation framework ( DeepSIL ) to include cooperative maneuvers, V2X communication, and edge-server environment models, and validates both approaches through comprehensive simulations and real-world experiments at the Ulm-Lehr pilot site. Results show that cooperative maneuvers substantially reduce crossing times and stops, with the optimization-based planner achieving greater efficiency but higher safety-criticality compared to the RL-based planner, which remains closer to non-cooperative baselines. The work demonstrates the practical feasibility of cooperative planning in real traffic and highlights trade-offs between efficiency gains and safety margins, informing future deployments with mixed traffic and vulnerable road users in mind.

Abstract

Cooperative maneuver planning promises to significantly improve traffic efficiency at unsignalized intersections by leveraging connected automated vehicles. Previous works on this topic have been mostly developed for completely automated traffic in a simple simulated environment. In contrast, our previously introduced planning approaches are specifically designed to handle real-world mixed traffic. The two methods are based on multi-scenario prediction and graph-based reinforcement learning, respectively. This is the first study to perform evaluations in a novel mixed traffic simulation framework as well as real-world drives with prototype connected automated vehicles in public traffic. The simulation features the same connected automated driving software stack as deployed on one of the automated vehicles. Our quantitative evaluations show that cooperative maneuver planning achieves a substantial reduction in crossing times and the number of stops. In a realistic environment with few automated vehicles, there are noticeable efficiency gains with only slightly increasing criticality metrics.

Figures (9)

• Figure 1: Cooperative maneuver at an unsignalized intersection. The CAV $\nu_3$ on the major road gives way to the turning CAV $\nu_1$ and the CAV $\nu_2$ on the subordinate road under consideration of the HDV (gray). We compare an optimizing planner and an RL-based planner for deployment on an edge server.
• Figure 2: Illustration of the observation input features listed in Table \ref{['tbl:agent_observation']}. Figure adapted from mertens_fast_2024.
• Figure 3: Graph-based input representation for mixed traffic of the RL-based planning module. A CAV's (yellow) turning intention is denoted by an arrow on its hood. Due to the unknown turning intention of the HDV $\nu_2$ (blue, denoted by '?'), it shares edges with both CAVs, although only the conflict with $\nu_5$ is inevitable.
• Figure 4: The initial positions of vehicles (yellow boxes) in the simulated scenarios. They are populated starting with the positions closest to the intersection. The evaluated trajectory segments are indicated by the black paths. The minor subordinate road approaches from the west.
• Figure 5: The simulative evaluation results in fully automated traffic for the optimization-based cooperative planner (Opt), the RL-based cooperative planner (RL), and non-cooperative (NC) CAVs. HDV denotes pure HDV traffic.