Abstract
Single-vehicle autonomy remains constrained by perception errors and uncoordinated maneuvers, resulting in avoidable collisions and throughput losses at intersections. Cooperative Driving Automation (CDA) offers substantial gains, yet fragmented toolchains hinder progress: industry develops closed systems, while academic studies rely on simplified simulators that lack realistic dynamics. We introduce OpenCDA-MARL, an open-source extension of OpenCDA that transforms the high-fidelity CARLA simulator into a unified benchmarking platform for autonomous intersection management (AIM). The framework provides a modular multi-agent reinforcement learning (MARL) API enabling reproducible training and comparison of cooperative policies under identical physical conditions. Building upon OpenCDA's integration and scenario managers, OpenCDA-MARL provides an additional MARL Adapter Interface layer, along with MARL scenario and environment managers for custom traffic conditions, performance tracking, baseline agents, and standard MARL algorithms, to support a comprehensive MARL evaluation suite in AIM. Key innovations include: (1) a 43-dimensional TTC-augmented observation model providing proactive collision awareness via physics-based trajectory prediction; (2) a cooperative yielding reward that incentivizes speed reduction when neighboring vehicles face collision risk; and (3) an LSTM-based multi-agent context encoder that captures spatiotemporal traffic interactions using distance-sorted agent sequences. OpenCDA-MARL includes rule-based, behavior, and vanilla agents, along with MARL baselines (MATD3, MADQN, MAPPO, MASAC) implemented under centralized training with decentralized execution. Evaluations using standardized metrics show that MASAC achieves perfect success at light traffic and maintains strong performance under heavy congestion (300 vph), substantially outperforming rule-based baselines. OpenCDA-MARL provides datasets, training scripts, and evaluation protocols to accelerate reproducible research toward safe and scalable CDA systems.