Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections

Coordinating autonomous vehicles at unsignalized intersections is a classic multi-agent challenge. Traditional multi-agent reinforcement learning (MARL) approaches often struggle with combinatorial action spaces, the need for privileged information (like future trajectories), or rigid, non-scalable agent designs. This paper asks whether a hierarchical architecture—decoupling high-level coordination strategy from low-level vehicle control—can improve performance and scalability in these complex traffic scenarios.

The authors introduce the Master-Agent Proto-plan System (MAPS). This architecture consists of two distinct components: a centralized Master agent and decentralized Worker agents. The Master observes the global traffic state and outputs a continuous, low-dimensional vector called a "proto-plan." This embedding serves as a global coordination signal. Each Worker agent receives this proto-plan alongside its local kinematic observations to determine its own actions. By using a continuous latent space for coordination, the system avoids the exponential growth of action spaces associated with traditional multi-agent systems and keeps communication overhead constant regardless of the number of vehicles.

MAPS was evaluated across 72 distinct intersection configurations in the HighwayEnv simulator. The system achieved collision-free navigation and reduced average travel time by 38% compared to state-of-the-art baselines. A key strength of the architecture is its generalization capability: a model trained with only three agents achieved a 94% success rate when deployed zero-shot to five-agent scenarios. This confirms that the learned proto-plans effectively capture transferable coordination strategies that do not require retraining when the fleet size changes.