DMAVA: Distributed Multi-Autonomous Vehicle Architecture Using Autoware

Simulating and validating coordination among multiple autonomous vehicles remains challenging, as many existing simulation architectures are limited to single-vehicle operation or rely on centralized control. This paper presents the Distributed Multi-Autonomous Vehicle Architecture (DMAVA), a simulation architecture that enables concurrent execution of multiple independent vehicle autonomy stacks distributed across multiple physical hosts within a shared simulation environment. Each vehicle operates its own complete autonomous driving stack while maintaining coordinated behavior through a data-centric communication layer. The proposed system integrates ROS 2 Humble, Autoware Universe, AWSIM Labs, and Zenoh to support high data accuracy and controllability during multi-vehicle simulation, enabling consistent perception, planning, and control behavior under distributed execution. Experiments conducted on multiple-host configurations demonstrate stable localization, reliable inter-host communication, and consistent closed-loop control under distributed execution. DMAVA also serves as a foundation for Multi-Vehicle Autonomous Valet Parking, demonstrating its extensibility toward higher-level cooperative autonomy. Demo videos and source code are available at: https://github.com/zubxxr/distributed-multi-autonomous-vehicle-architecture.

💡 Research Summary

The paper introduces DMAVA (Distributed Multi‑Autonomous Vehicle Architecture), a novel simulation framework that enables concurrent execution of multiple independent autonomous‑driving stacks across several physical hosts while sharing a common virtual environment. Existing simulators either focus on macroscopic traffic (e.g., SUMO, VISSIM) or high‑fidelity vehicle dynamics (e.g., CarSim, SCANeR) but lack native support for ROS 2‑based distributed autonomy. DMAVA bridges this gap by tightly integrating four open‑source components: ROS 2 Humble, Autoware Universe, AWSIM Labs (a Unity‑based sensor and physics simulator), and Zenoh (a lightweight data‑centric publish‑subscribe middleware).

Key architectural elements:

1. AWSIM Labs provides realistic LiDAR, camera, and IMU streams and handles vehicle physics. The authors modify the simulator to bypass the proprietary Vehicle Physics Pro (VPP) dependency, duplicating vehicle prefabs so that multiple vehicles can coexist in a single scene, each assigned a unique ROS 2 namespace.
2. Autoware runs as a full stack (perception, localization, planning, control) for each vehicle. Because Autoware expects a flat namespace, DMAVA uses Zenoh to strip and re‑apply vehicle‑specific prefixes on the fly, preserving isolation without altering Autoware code.
3. Zenoh operates in router mode on the simulation container and client mode on each vehicle‑autonomy container. JSON5 configuration files define namespace routing, topic filtering, and QoS, allowing high‑bandwidth sensor topics (e.g., raw images) to be excluded from inter‑host traffic, thereby reducing network load and latency.
4. Mapping & Localization: OpenStreetMap data are converted to Lanelet2 and 3D point‑cloud maps. The authors refine the map in Blender, adding planar walls to improve NDT scan‑matching stability near vehicle spawn points. Host‑specific NDT parameters balance accuracy against computational constraints.

Experimental evaluation comprises two‑host and three‑host deployments. In the two‑host setup, the Nitro PC (i7‑12700F, RTX 3060) hosts the simulation container and one vehicle’s autonomy stack, while a second laptop (Ryzen 7 4800HS, RTX 2060) runs the other autonomy stack. Results show localization errors below 0.15 m, control command latency under 30 ms, and stable closed‑loop operation for several minutes. Extending to three hosts (adding a third laptop with RTX 4050) demonstrates linear scalability: CPU/GPU utilization stays under 70 % and inter‑host latency increases by less than 5 ms. The authors also implement a multi‑vehicle autonomous valet‑parking scenario, where each vehicle independently plans and executes a parking maneuver while the shared simulator coordinates their motion via Zenoh. All vehicles reach their designated spots without collisions, confirming the framework’s suitability for cooperative autonomy research.

Contributions:

• First ROS 2‑native, fully distributed multi‑vehicle simulation architecture within the Autoware ecosystem.
• Zenoh‑based namespace routing that eliminates topic collisions and efficiently filters high‑bandwidth streams.
• Modification of AWSIM Labs to support multiple vehicles without requiring a commercial VPP license.
• Practical map‑preprocessing and NDT tuning to achieve robust localization in simulation.
• Demonstration of a higher‑level cooperative task (autonomous valet parking) as a proof‑of‑concept.

Limitations include the per‑vehicle one‑Autoware‑instance design, which may become a bottleneck as the number of vehicles scales dramatically, and reliance on the Unity free version, which offers less precise physics than the paid VPP module. Future work is outlined to incorporate direct V2V communication over Zenoh, integrate hardware‑in‑the‑loop components, and explore cloud‑based auto‑scaling for thousands of simulated vehicles.

In summary, DMAVA provides an open, scalable, and high‑fidelity testbed for multi‑autonomous‑vehicle research, enabling developers to evaluate perception, planning, and coordination algorithms under realistic sensor conditions while distributing computational load across multiple machines.