Scaled-Down Multi-Vehicle Cooperative Experimental Platform

This project presents a scaled-down physical vehicle experimental platform designed to validate multi-vehicle cooperative decision-making algorithms, achieving a cyber-physical closed loop characterized by centralized perception and decision-making alongside distributed low-level execution. Built upon the Robot Operating System (ROS), the hardware architecture consists of an Ackermann wheeled robot cluster, a central decision server acting as an edge computing node, and a structured physical testbed equipped with high-precision localization. The overarching system architecture highly aligns with the advanced development trend of “Vehicle-Road-Cloud” integration in intelligent transportation, providing an ideal testing environment for evaluating the effectiveness of cutting-edge reinforcement learning algorithms under real-world physical constraints.

Physical environment layout and hierarchical software-hardware communication architecture of the experimental platform. Through high-frequency wireless interactions between onboard controllers and the edge server, the platform accurately replicates real-world "Vehicle-Road-Cloud" cooperative scenarios.

In terms of experimental design, this project focuses on the classic “zipper merge” conflict scenario, which features high game-theoretic complexity, to test the algorithm’s capability in resolving conflicts between individual optimums and collective swarm efficiency. To bridge the “Sim-to-Real” gap from pure virtual simulation to physical world deployment, the software system innovatively introduces an adjoint state projection mechanism, resolving perception overlaps and safety hazards caused by the continuity of lane-changing maneuvers in physical space. Concurrently, the low-level controllers fully account for communication latency, sensor noise, and authentic vehicle dynamics constraints, ensuring that high-level semantic decisions are executed precisely and smoothly by the robots.

Physical execution snapshots of the Topology-Enhanced Multi-Agent Reinforcement Learning (TPE-MARL)(Han et al., 2026) algorithm operating in a zipper merge scenario.

Physical experimental results demonstrate that the vehicle cluster, driven by the core cooperative algorithm, successfully exhibits proactive, high-level emergent intelligent behaviors in reality. In comparative tests, this cooperative strategy not only induces vehicles to yield proactively to create global lane-changing gaps but also significantly reduces the total system task time, achieving substantial cooperative efficiency gains. System performance monitoring further reveals that the trajectory tracking error during physical execution is minimal, and the end-to-end decision latency falls well below control cycle requirements, fully proving the engineering feasibility and robust stability of this multi-agent decision-making architecture for real-world intelligent transportation system deployments.

Spatiotemporal trajectory comparison under different strategies. Vehicles governed by the cooperative strategy display smoother and more efficient spatiotemporal interaction characteristics.

Decentralized multi-agent cooperative decision-making in continuous environments is fundamentally bottlenecked by the curse of dimensionality, where undirected exploration typically converges to conservative local optima. We propose Topology-Enhanced Multi-Agent Reinforcement Learning (TPE-MARL) to reformulate multi-agent exploration as a structured topological traversal. We introduce the Game Topology Tensor, utilizing locality-sensitive hashing to project the continuous physical manifold into a discrete quotient space. This abstraction operates as an adversarial Information Bottleneck, decoupling strategic coordination intents from environmental noise. Within this space, a dual intrinsic reward mechanism drives exploration: a novelty reward maximizes the marginal entropy of visited topologies, while a collaboration reward, optimized via a variational Evidence Lower Bound (ELBO), minimizes conditional entropy to exploit cooperative joint configurations. Evaluations demonstrate that TPE-MARL achieves near-optimal decision distributions, closely approximating the theoretical bounds established by a Monte Carlo Tree Search (MCTS) oracle. Furthermore, physical testbed experiments validate the framework’s zero-shot out-of-distribution (OOD) generalization. Supported by a spatial relaxation mechanism, the learned representations reliably execute dynamic negotiations, such as cooperative zipper-merging, exhibiting inherent robustness against real-world covariate shifts and actuation latencies.

References

2026