eCAV: An Edge-Assisted Evaluation Platform for Connected Autonomous Vehicles

Abstract: As autonomous vehicles edge closer to widespread adoption, enhancing road safety through collision avoidance and minimization of collateral damage becomes imperative. Vehicle-to-everything (V2X) technologies, which include vehicle-to-vehicle (V2V), vehicle-to-infrastructure (V2I), and vehicle-to-cloud (V2C), are being proposed as mechanisms to achieve this safety improvement. Simulation-based testing is crucial for early-stage evaluation of Connected Autonomous Vehicle (CAV) control systems, offering a safer and more cost-effective alternative to real-world tests. However, simulating large 3D environments with many complex single- and multi-vehicle sensors and controllers is computationally intensive. There is currently no evaluation framework that can effectively evaluate realistic scenarios involving large numbers of autonomous vehicles. We propose eCAV -- an efficient, modular, and scalable evaluation platform to facilitate both functional validation of algorithmic approaches to increasing road safety, as well as performance prediction of algorithms of various V2X technologies, including a futuristic Vehicle-to-Edge control plane and correspondingly designed control algorithms. eCAV can model up to 256 vehicles running individual control algorithms without perception enabled, which is $8\times$ more vehicles than what is possible with state-of-the-art alternatives.

• The paper presents eCAV, a scalable platform that fully simulates the CAV autonomy stack with integrated edge-node support.
• It employs parallel containerization and realistic V2X network emulation to accurately benchmark control algorithms at unprecedented scales.
• Empirical results demonstrate evaluation with up to 256 vehicles, minimal communication overhead, and enhanced simulation fidelity.

eCAV: An Edge-Assisted Evaluation Platform for Connected Autonomous Vehicles

The paper introduces eCAV, an evaluation platform designed for large-scale simulation and assessment of Connected Autonomous Vehicles (CAVs) in environments featuring advanced edge-computing support. The primary contribution is a scalable, modular, and extensible system that enables realistic evaluation of CAV control algorithms, encompassing both traditional in-vehicle autonomy and futuristic edge-assisted paradigms.

Motivation and Background

Thorough validation of CAV algorithms via simulation is essential due to cost, safety, and scalability limitations of real-world testing. Existing simulators—such as CARLA, OpenCDA, and AutoCastSim—are inadequate for several reasons: they are unable to scale to hundreds of vehicles, lack comprehensive edge computation modeling, and often prioritize performance at the expense of simulation fidelity. Notably, state-of-the-art solutions either (a) reduce realism by shortcutting parts of the autonomy stack, or (b) cannot evaluate collaborative edge-based methods at scale.

eCAV is positioned as a response to these deficiencies. Key differentiators include:

• Full autonomy stack simulation for each vehicle, including perception, planning, and control.
• Modular edge-node interface for pluggable, user-defined multi-vehicle coordination algorithms.
• Network emulation capabilities for realistic modeling of V2X communication characteristics.
• Distributed execution across heterogeneous compute resources and multi-node clusters.

Architectural Overview

The architecture decomposes simulation into independent, containerized actors: vehicle clients, an environment simulator (CARLA), a simulation manager, an edge node, a V2X communication collector, and a network emulator. Major design features include:

• Parallelization: Each vehicle runs in a dedicated container, with computation decoupled from other agents, allowing effective scaling within and across machines. This tackles Python GIL constraints via process-level isolation and enables true multi-core and multi-node utilization.
• Edge Node Integration: Pluggable interfaces support deploying arbitrary edge-resident control logic, facilitating experimentation with centralized coordination, perception fusion, or global path planning paradigms.
• Communication Architecture: Event-driven, asynchronous, and hardware-agnostic push-pull messaging is implemented via gRPC and protobuf, ensuring parallelism without bottlenecking on synchronization.
• Metrics Aggregation: Non-blocking, distributed metric collection—batched and submitted only after simulation steps—ensures that runtime scalability is not compromised by data analytics.

Implementation Considerations

Containers manage vehicle client isolation, with nVidia's Docker runtime used for perception-enabled scenarios that require GPU acceleration. The system coordinates tightly with CARLA in synchronous mode, maintaining simulation determinism.

Distribution and Heterogeneity are managed by deploying vehicle clients and simulation components across clusters, making resource allocation flexible and mapping well onto cloud or high-performance compute environments.

Extensibility is provided through YAML-based scenario, vehicle, and edge algorithm configuration, and a Python scripting interface to define custom coordination policies (e.g., how and when the edge broadcasts waypoints).

Performance and Empirical Results

Strong empirical results are demonstrated:

• Scalability: Simulations without perception run up to 256 vehicles (8x over prior art); with perception, up to 64 vehicles are supported at 800ms step times (4x more and 1.5x faster than OpenCDA).
• Resource Utilization: Multi-GPU and multi-node distribution is efficiently exploited. The overhead introduced by eCAV for network communication and barrier synchronization is consistently small (<50ms at the 99th percentile, even with 256 vehicles).
• Fidelity: Vehicle-level localization traces and behavior metrics are shown to be statistically identical to OpenCDA's output, affirming that scalability is not achieved at the expense of simulation fidelity.
• Usability: Integration with arbitrary edge algorithms (e.g., A*-based multi-agent trajectory planning and velocity clustering via k-means) is straightforward. Metrics including control latency, deviation from target velocity, and safety violations are automatically recorded.

Claims and Implications

eCAV claims:

• The first platform supporting large-scale, perception-enabled, edge-assisted CAV evaluation with realistic network modeling and independent control stacks.
• Linear scaling with compute resources and extensibility to new coordination scenarios and control paradigms.
• Seamless integration of edge-control latency into closed-loop AV simulation, including agent fallback to local data in the face of stale edge input.

The practical implications are substantial: eCAV enables the CAV research community to (a) quantitatively assess the effects of edge-assistance under realistic latency and failure scenarios, (b) benchmark different collaborative control algorithms at unprecedented scales, and (c) accelerate iteration by eliminating trade-offs between performance and simulation realism.

Theoretical and Practical Outlook

eCAV represents an inflection point for simulation-driven CAV research, aligning with the trajectory toward edge-centered vehicular autonomy and heterogenous urban traffic. By decoupling algorithm definition, system architecture, and deployment topology, the platform facilitates the co-design of perception, planning, V2X protocols, and control under future-relevant failure and heterogeneity regimes.

Potential future research enabled by eCAV includes:

• Evaluation of multi-edge architectures (e.g., metropolitan-scale edge clouds), where load balancing and handoff protocols can be studied with hundreds of AVs across network partitions.
• Investigation of space-sharing strategies for perception to further boost simulation scale (e.g., via shared sensory caches or sensor model abstraction).
• Large-scale, closed-loop validation of safety metrics under adversarial traffic, realistic dropout, and attack models.
• Co-optimization of edge-control policies and network infrastructure for latency, bandwidth, and compute efficiency.

Conclusion

eCAV addresses a critical bottleneck in evaluating next-generation edge-assisted CAV systems, delivering an open, scalable, and configurable platform that sets a new standard for simulation fidelity, scale, and flexibility. By supporting large-scale benchmarking of distributed and collaborative autonomy algorithms—not merely theoretical but under constraints and failure patterns of real-world V2X networks—it represents a key enabler for both the academic and industrial CAV communities.