Route-Time Conflict Maps for Cooperative End-to-End Autonomous Driving

Haohua Que, Mingkai Liu, Tianle Zhu, Zhipeng Bao, Yiping Li, Handong YaoarXiv 2026 · arXiv

TL;DR

When2Go is a cooperative end-to-end autonomous driving system that moves beyond simple perception sharing to intent sharing. By representing the future as a 2D "route-time conflict map" and using a decentralized "order gate" to agree on passing priority, it achieves state-of-the-art performance across multiple benchmarks, virtually eliminating the deadlocks and collisions common in multi-vehicle interactions.

What Problem Does This Paper Address?

Most current cooperative driving research focuses on "Class A" automation: vehicles share what they see (perception) but not what they plan to do (intent). When multiple autonomous vehicles (AVs) meet at an intersection or a merge point, this lack of shared intent leads to two major failure modes:

  1. Polite Deadlocks: Every vehicle yields, and no one moves.
  2. Collisions: Multiple vehicles assume they have the right of way and move simultaneously.

While some systems try to solve this with multi-round language-based negotiation, these methods are often slow, computationally expensive, and brittle under noisy communication. The authors of this paper sought a way for vehicles to reach a decentralized agreement on passing order instantly and robustly.

Methods

The core of the When2Go system is a shift toward Class B (intent-sharing) cooperative driving.

  • Route-Time Conflict Map: Instead of planning on a standard 2D Bird's Eye View (BEV) map, When2Go projects the world onto a 2D grid where one axis is the distance along the vehicle's route ($s$) and the other is future time ($t$). This "curtain" representation (see Figure 3 below) makes it immediately obvious if a point on the route will be occupied by another vehicle at a specific future time.
  • Order Gate: To decide who goes first, vehicles broadcast an "intended trajectory"—the path they would take if the road were empty. A small, antisymmetric precedence network (the "order gate") compares two trajectories and decides the passing order. Because the network is antisymmetric, any two vehicles evaluating the same messages will always reach the same conclusion, ensuring agreement without back-and-forth negotiation.
  • End-to-End Architecture: The system uses standard LiDAR encoders and BEV backbones to process local and shared perception data into "instance tokens." These tokens, along with shared plans, are fused and fed into the route-time map and planner.
  • Smart Supervision: The model is trained to imitate a rule-based expert. Crucially, the authors use "relabelling" for stops the model can't explain (e.g., stopping for a red light that isn't in the LiDAR input). This prevents the model from learning to stop for "ghost" causes, which would otherwise lead to overly timid behavior.
The system architecture of When2Go, showing the flow from LiDAR/messages to the route-time conflict map and final plan.

Datasets

The model was trained on released recordings from the V2Xverse dataset (Town01-04 and 06). Evaluation was conducted in closed-loop simulations across three major CARLA-based benchmarks:

  • InterDrive: Focuses on complex multi-vehicle interactions at junctions and merges.
  • V2Xverse: A public protocol featuring a single controlled vehicle and a roadside unit (RSU) in dense traffic.
  • MDrive: A benchmark featuring scenarios replayed from real-world logs.

Key Insights and Ideas

  • Dimensionality Reduction for Planning: The "aha moment" is realizing that for a vehicle following a fixed route, the only decision that matters is when to be at a certain where. Mapping the 4D world (3D space + time) into a 2D route-time map (1D route + time) simplifies the planning problem into a simple occupancy check.
  • Zero-Shot Agreement: By using an antisymmetric precedence network, the "order gate" achieves decentralized agreement in a single forward pass. This eliminates the latency and complexity of multi-round negotiation.
  • Temporal Asymmetry: The authors recognized that a learner with a 2-second planning horizon shouldn't try to imitate an expert reacting to events 5 seconds away. Relabelling these "unexplained" stops is vital for training a model that is decisive rather than paralyzed by uncertainty.
A visualization of the route-time conflict map. The ego vehicle's route is unrolled into a 2D grid of position vs. time, where purple blocks represent predicted occupancy by other vehicles.

Key Results

When2Go set new state-of-the-art (SOTA) marks on all three benchmarks, significantly outperforming previous methods:

  • InterDrive: Achieved a driving score of 97.8 (without traffic) and 90.7 (with traffic), surpassing the previous best of 90.1 and 78.9, respectively.
  • V2Xverse: Reached a score of 70.0, well above the best of sixteen rerun methods (59.1).
  • MDrive: Scored 88.2, compared to the best rerun of 85.1.
  • Safety: On the MDrive benchmark, 95.0% of When2Go's vehicles were collision-free, compared to 87.7% for the safest baseline.

Limitations and Future Work

While highly effective, the authors note several areas for improvement:

  • Sensing Modalities: The current system relies solely on LiDAR. It lacks inputs for traffic lights and stop signs, which are currently handled through the relabelling training trick.
  • Fixed Routing: The planner follows a pre-defined route and does not yet perform its own autonomous lane changes or route planning.
  • Communication Reliability: While the paper tests "imperfect channels" (latency and message loss), the core "order gate" logic assumes vehicles can eventually hear each other to reach an agreement. Future work could explore even more robust decentralized coordination under extreme communication failures.

Summary updated

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%).
BibTeX
@misc{que2026route,
  author = {Haohua Que and Mingkai Liu and Tianle Zhu and Zhipeng Bao and Yiping Li and Handong Yao},
  title = {{Route-Time Conflict Maps for Cooperative End-to-End Autonomous Driving}},
  year = {2026},
  eprint = {2611.13494},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13494},
}
Figure