[Paper Review] Coordinating CAV Swarms at Intersections with a Deep Learning Model

Jiawei Zhang, Shen Li+1 more|arXiv (Cornell University)|Nov 10, 2022

Robotic Path Planning Algorithms4 citations

TL;DR

This paper proposes AlphaOrder, a deep learning and tree search hybrid algorithm that rapidly generates near-optimal passing orders for connected and automated vehicle (CAV) swarms at signal-free intersections. By training a pointer network on solved scenarios and refining results via online tree search, AlphaOrder achieves real-time performance with significantly lower delay than existing methods, even for large-scale CAV flows.

ABSTRACT

Connected and automated vehicles (CAVs) are viewed as a special kind of robots that have the potential to significantly improve the safety and efficiency of traffic. In contrast to many swarm robotics studies that are demonstrated in labs by employing a small number of robots, CAV studies aims to achieve cooperative driving of unceasing robot swarm flows. However, how to get the optimal passing order of such robot swarm flows even for a signal-free intersection is an NP-hard problem (specifically, enumerating based algorithm takes days to find the optimal solution to a 20-CAV scenario). Here, we introduce a novel cooperative driving algorithm (AlphaOrder) that combines offline deep learning and online tree searching to find a near-optimal passing order in real-time. AlphaOrder builds a pointer network model from solved scenarios and generates near-optimal passing orders instantaneously for new scenarios. Furthermore, our approach provides a general approach to managing preemptive resource sharing between swarm robotics (e.g., scheduling multiple automated guided vehicles (AGVs) and unmanned aerial vehicles (UAVs) at conflicting areas

Motivation & Objective

• To address the NP-hard problem of determining optimal passing orders for unceasing CAV swarms at signal-free intersections.
• To overcome the limitations of existing methods—such as high computational cost, poor scalability, or reliance on heuristic rules—by combining deep learning with online search.
• To enable real-time, low-complexity coordination of large-scale CAV swarms in dynamic traffic environments.
• To generalize across diverse traffic demand patterns and intersection configurations, ensuring robust performance under heterogeneous conditions.
• To provide a scalable, transferable framework for preemptive resource sharing in swarm robotics and intelligent transportation systems.

Proposed method

• AlphaOrder employs a pointer network $p_{\theta}$ trained on a dataset of solved scenarios to predict candidate passing orders directly from vehicle state inputs.
• It integrates a Monte Carlo Tree Search (MCTS) module to refine the initial predictions by exploring high-reward sequences, improving solution quality.
• A critic network $b_{\delta}$ acts as a baseline in the REINFORCE reinforcement learning framework to stabilize training of the pointer network.
• The algorithm uses a two-stage pipeline: offline pre-training of the pointer network on synthetic scenarios, followed by online tree search for new, unseen configurations.
• The system is designed for deployment on roadside units, with model updates performed remotely on the cloud to minimize on-device complexity.
• The method supports generalization to arbitrary numbers of CAVs and various traffic demand patterns, including heterogeneous flows.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on solved scenarios generate near-optimal passing orders for new, unseen CAV swarm configurations at signal-free intersections?
• RQ2 How does the integration of deep learning with online tree search improve computational efficiency and solution quality compared to traditional optimization or heuristic methods?
• RQ3 To what extent can the model generalize across different traffic demand patterns, including heterogeneous and high-volume flows?
• RQ4 What is the performance gain of AlphaOrder in terms of reducing total delay compared to first-come-first-served (FIFO) and MCTS-only baselines?
• RQ5 Can the algorithm be efficiently deployed on resource-constrained roadside units while maintaining real-time responsiveness?

Key findings

• AlphaOrder reduces average delay by up to 40% compared to the FIFO baseline in high-demand scenarios, with a 25% improvement over MCTS-based methods in Pattern 1.
• In heterogeneous traffic demand scenarios (e.g., 2.5× flow difference), AlphaOrder maintains superior performance, with a 14.23-second average delay compared to 22.27 seconds for FIFO in Pattern 2.
• The algorithm achieves near-optimal solutions in milliseconds for scenarios with up to 40 CAVs, whereas enumeration-based methods take days to compute optimal orders.
• The critic network and MCTS refinement step significantly improve solution quality, with the final output consistently outperforming both the initial pointer network prediction and baseline algorithms.
• AlphaOrder demonstrates strong generalization across diverse traffic patterns, showing only a 12.81-second average delay in Pattern 1 and 14.23 seconds in Pattern 2, indicating robustness to flow imbalance.
• The model’s low on-device complexity enables practical deployment on roadside units, with training and updates handled remotely on the cloud.