Zero-Collision Intersection Navigation via Compact Latent Coordination

Researchers at Ben-Gurion University developed MAPS, a hierarchical deep reinforcement learning architecture that enables autonomous vehicles to navigate unsignalized intersections without a single collision. The system uses a compact 4-dimensional "proto-plan" embedding to coordinate multiple vehicles, reducing travel time by 38% compared to existing methods while scaling seamlessly to larger fleets without retraining.

What the Researchers Built

MAPS (Multi-Agent Proto-Plan System) is a hierarchical deep reinforcement learning architecture purpose-built for coordinating multiple autonomous vehicles through unsignalized intersections—junctions without traffic lights or stop signs. The system splits decision-making into two levels: a centralized Master agent that sees the full global state (positions and velocities of all nearby vehicles) and outputs a continuous “proto-plan” vector, and decentralized Worker agents that receive this embedding alongside their own local observations to execute low-level driving commands.

The proto-plan replaces discrete, predefined coordination commands (like “vehicle 1 yields, vehicle 2 proceeds”) with a dense 4-dimensional vector learned end-to-end. This avoids the combinatorial explosion of action spaces as fleet size grows—complexity stays linear rather than exponential. The Master is trained using a reward structure that encourages safety and efficiency: each Worker’s reward is based on its own traversal, but the team optimizes the minimum reward across agents (a maximin approach) to prevent any vehicle from being left behind.

The team evaluated MAPS across 72 intersection configurations that vary approach direction, turning intention, and initial distance. Crucially, the model trains on scenarios with only three active vehicles but can be deployed to five-vehicle intersections without any fine-tuning—achieving 94% success in zero-shot transfer. The system uses only readily available kinematic state (positions and velocities) and requires no privileged information like future trajectories, expert demonstrations, or rule-based safety layers.

Key Results

MAPS was evaluated against two recent multi-agent coordination baselines: VN-MADDPG (value decomposition with demonstrations) and QMIXwD (actor-critic with adaptive exploration). The results show a clear margin in both safety and efficiency.

The safety advantage extends beyond evaluation: MAPS incurred only 21 collisions across 900 training episodes, an 84–85% reduction relative to the baselines (132 and 147). A two-proportion z-test confirmed statistical significance at p < 0.01. The average travel time of 7.8 steps represents a 38% improvement over the best-performing baseline. In zero-shot testing from 3 to 5 agents, MAPS maintained a 94% success rate over 100 episodes, with the few failures occurring in high-density turning scenarios.

How It Works

MAPS employs a hierarchical reinforcement learning structure with two distinct policy levels:

Master Agent (Centralized): The Master takes as input a concatenated global state vector containing the kinematic information (positions and velocities) of all vehicles in the scene. The system pads slots for up to a fixed maximum number of vehicles—when fewer are active, the remaining slots are zero-padded. The Master’s policy outputs a proto-plan embedding, a continuous vector of 4 dimensions that encodes the intended coordination strategy. Unlike discrete high-level commands, this dense representation allows the strategy to be smoothly adapted to the exact traffic situation.

Worker Agents (Decentralized): Each vehicle is controlled by a Worker agent that receives both its own local observation (ego-centric view) and the shared proto-plan from the Master. All Worker agents share parameters, which reduces the total number of trainable parameters and forces the policy to learn generalizable driving behaviors. Because Workers operate on local observations with a unified controller, the same policy can handle vehicles at different positions and turning intentions.

Reward Structure: Each Worker receives a reward that balances progress toward its goal (speed and heading alignment) against collision penalties and proximity to other vehicles. The team’s objective is to maximize the minimum reward across all Workers—the maximin approach—ensuring that no vehicle is left stranded or forced into an unsafe maneuver.

Training Protocol: MAPS is trained using the HighwayEnv simulator with simplified kinematics. The training uses only 3 active agents across 72 intersection configurations (18 base layouts × 4 rotations). After training, the same model is directly deployed to 5-agent scenarios with no fine-tuning. The key to this generalization lies in three architectural choices: the proto-plan embedding is agent-count agnostic; parameter sharing across Workers works for any number of vehicles; and the maximin reward incentivizes strategies that don’t assume a fixed agent count.

Why This Matters for Robotics

MAPS demonstrates that effective multi-vehicle coordination at unsignalized intersections can be achieved with an extremely compact communication signal—just 4 floating-point numbers per time step. This is critical for real-world deployment in vehicle-to-everything (V2X) systems, where bandwidth is limited and latency matters. The proto-plan embedding fits easily into existing communication protocols.

For robotics beyond road vehicles, this approach has direct parallels. Warehouse robots navigating narrow aisles, autonomous forklifts at loading docks, and even multi-robot assembly cells face similar coordination challenges—multiple agents must cross paths without collisions and without a central traffic controller. The ability to train on small team sizes and deploy to larger fleets without retraining (94% zero-shot success) slashes the cost of deploying multi-robot systems. The fact that MAPS requires no map data, no expert demonstrations, and no future trajectory prediction makes it practical for dynamic, unstructured environments.

For fleet operators, this could mean simpler programming: define the intersection geometry, let the system learn to coordinate, and scale up by adding more robots. The architecture’s reliance on only kinematic state (positions and velocities) means it can be retrofitted to existing robots with basic odometry and communication modules. Check out warehouse robots on Robot Overflow and used industrial robots to see platforms that could benefit from this coordination method.

Limitations and Open Questions

MAPS was evaluated only in the HighwayEnv kinematic simulator, which abstracts away perception noise, sensor latency, and detailed vehicle dynamics. Real intersections involve uncertainty in state estimation, communication delays, and complex tire dynamics that could break the simplified assumptions. Validation in higher-fidelity simulators like SUMO or CARLA is needed.

The system assumes that the Master has access to the exact positions and velocities of all vehicles—an assumption that may not hold in real-world V2X deployments where sensor fusion and communication dropouts are common. The proto-plan is a single-step embedding; extending it to multi-step plans could amortize communication costs but would require new temporal reasoning mechanisms. Finally, the current work uses homogeneous agents; heterogeneous fleets (cars, trucks, pedestrians) would require vehicle-class-specific Worker modules.

Frequently Asked Questions

What is a proto-plan embedding? A proto-plan is a dense, continuous vector (4 dimensions) output by a Master agent that encodes a coordination strategy for all vehicles at an intersection. It replaces discrete commands like “vehicle 1 yields” with a learned representation that smoothly adapts to traffic conditions.

Does MAPS require communication between vehicles? Yes, but only a small amount: the Master broadcasts the 4-dimensional proto-plan to all Workers each time step. This is much more bandwidth-efficient than sharing raw observations or multiple discrete commands.

How does MAPS generalize to new intersection layouts? The model was trained on 72 configurations (18 base layouts rotated through 4 orientations). The hierarchical architecture and shared Worker parameters allow it to handle any approach direction and turning intention without retraining.

Can MAPS coordinate more than 5 vehicles? The paper tested zero-shot transfer from 3 to 5 agents with 94% success. The architecture can theoretically handle more agents because the proto-plan dimension is fixed and Workers are instance-agnostic, but performance at larger sizes was not evaluated.

Conclusion

MAPS introduces a practical, scalable solution for multi-vehicle coordination at unsignalized intersections using a tiny 4-dimensional coordination signal. With zero collisions, 38% faster travel, and robust zero-shot transfer to larger fleets, this hierarchical DRL approach could simplify deployment of autonomous vehicles and multi-robot systems in constrained environments.