Route-Time Conflict Maps for Cooperative End-to-End Autonomous Driving
Abstract
Existing cooperative end-to-end driving works at Class A (status-sharing) of the SAE J3216 classes of cooperative driving automation (CDA): vehicles exchange what they perceive, not what they intend, so each plans alone when to go, which limits cooperation and leads to deadlock when both vehicles yield or to collisions when both go. We present When2Go, a cooperative end-to-end driving system that learns when to go by sharing intent under CDA Class B or higher. Each vehicle broadcasts its LiDAR detections with forecast motion as instance tokens, together with its plan and intended trajectory. The receiver draws the fused detections, forecasts and partner plans into a route-time conflict map, the occupancy along its route over the planning horizon, on which a planner scores candidate speed profiles. Where two intended trajectories conflict, an order gate, a precedence network distilled from a right-of-way rule, decides who goes first; both vehicles evaluate it on the same messages and so reach the same order. When2Go imitates the rule-based expert of V2Xverse, which sees the simulator's ground truth; stops that When2Go cannot observe within its 2 s planning horizon are relabelled, so that it does not learn to wait without an observable cause. For our runs only, the benchmarks' PID controller steers by When2Go's waypoints instead of the route, and its forced push of long-waiting vehicles is off. With one model, When2Go sets the state of the art on three closed-loop CARLA benchmarks: 97.8 and 90.7 driving score on InterDrive without and with traffic (best published 90.1 and 78.9), 70.0 on the public V2Xverse protocol (best of sixteen rerun methods 59.1) and 88.2 on MDrive (best rerun 85.1), with 95.0% of its vehicles collision-free (87.7%).
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper introduces a cooperative end-to-end driving system that leverages intent sharing and a route-time conflict map to resolve traffic conflicts, achieving state-of-the-art results on major benchmarks despite relying on modified evaluation controllers. The technical contribution is significant in advancing the state of cooperative driving automation, offering a robust framework for multi-agent planning that addresses the deadlock and collision issues inherent in status-sharing-only systems.
Methodology
The paper proposes "When2Go," a cooperative end-to-end driving system that moves beyond Class A (status-sharing) to Class B (intent-sharing) cooperative driving automation. The core methodological contribution is the "route-time conflict map," a representation that fuses local LiDAR detections, motion forecasts, and partner vehicle plans to predict occupancy along the ego vehicle's route over a planning horizon. A key innovation is the "order gate," a precedence network distilled from right-of-way rules that resolves conflicts by determining which vehicle has priority. This allows both vehicles to evaluate the same messages and reach a consistent decision on who goes first, mitigating deadlocks and collisions. The training strategy involves imitating a rule-based expert from the V2Xverse simulator, with a specific relabeling mechanism for stops that are not observable within the 2-second planning horizon to prevent the model from learning to wait without cause.
Results & evaluation
The experimental evaluation is rigorous, utilizing three major closed-loop CARLA benchmarks: InterDrive (with and without traffic), V2Xverse, and MDrive. The results are strong, claiming state-of-the-art performance on all three: 97.8/90.7 on InterDrive, 70.0 on V2Xverse, and 88.2 on MDrive. The collision-free rate is reported at 95.0%. However, the paper admits to a significant experimental caveat: for these runs, the benchmarks' standard PID controller was replaced with one that steers by When2Go's waypoints, and a "forced push" mechanism for long-waiting vehicles was disabled. This suggests the results may be partially dependent on specific controller modifications rather than purely the planning model's superiority in a standard setup.
Reproducibility
Reproducibility is moderate. The paper describes the architecture and training data sources (V2Xverse expert) clearly. However, the specific modifications to the evaluation protocol (PID controller changes, disabling forced push) are critical details that must be replicated exactly to achieve the reported scores. Without access to the specific code for these modified controllers, independent verification of the SOTA claims is difficult. The reliance on a specific simulator (CARLA) and a specific expert (V2Xverse) limits the immediate generalizability of the exact implementation.
Limitations
The primary limitation is the dependence on the modified evaluation protocol. The claim of SOTA is contingent on using a non-standard controller configuration. Additionally, the system relies on V2X communication (Class B CDA), which assumes reliable, low-latency connectivity between vehicles, a condition not always met in real-world deployments. The "order gate" is distilled from rule-based right-of-way, which may not capture all nuanced human driving behaviors or ambiguous traffic scenarios.
Broader impact
The work has significant implications for the development of cooperative autonomous driving systems. By demonstrating that sharing intent (plans) rather than just perception (status) leads to superior performance in conflict resolution, it provides a strong argument for higher-level V2X protocols. The route-time conflict map is a useful representation that could be adopted by other planning systems. The findings contribute to the broader goal of safe and efficient multi-agent traffic flow, potentially reducing congestion and accidents in mixed-traffic environments. The paper introduces a cooperative end-to-end driving system that leverages intent sharing and a route-time conflict map to resolve traffic conflicts, achieving state-of-the-art results on major benchmarks despite relying on modified evaluation controllers. The technical contribution is significant in advancing the state of cooperative driving automation, offering a robust framework for multi-agent planning that addresses the deadlock and collision issues inherent in status-sharing-only systems.
Author of this paper? Send a correction.