Distributed Real-Time Vehicle Control for Emergency Vehicle Transit: A Scalable Cooperative Method

This paper proposes a scalable, distributed vehicle control method that enables emergency vehicle transit by allowing ordinary vehicles to make real-time, approximately optimal decisions using only local information, thereby overcoming the computational and scalability limitations of existing centralized and learning-based approaches while guaranteeing safety through a novel conflict resolution mechanism.

The Gist

Imagine a busy highway as a giant, flowing river of cars. Suddenly, an ambulance (an Emergency Vehicle or EMV) needs to get to a hospital as fast as possible. In the real world, this usually means the ambulance flashes its lights and siren, and the other cars (Ordinary Vehicles or OVs) panic, slam on their brakes, or swerve chaotically to get out of the way. This creates a traffic jam, delays the ambulance, and makes everyone else late.

This paper proposes a smarter, calmer way to handle this situation using a new system called SDVC (Scalable Distributed Vehicle Control).

Here is the breakdown of the problem and their solution, using simple analogies:

The Problem: The "Conductor" vs. The "Chaos"

Previous methods tried to solve this in two ways, both of which had big flaws:

1. The "Super-Conductor" (Centralized Solvers): Imagine one giant brain in a control tower trying to tell every single car on the highway exactly what to do, second by second.
   • The Flaw: If there are only 20 cars, the brain can do the math quickly. But if there are 400 cars, the brain gets overwhelmed. It takes too long to calculate, and by the time it figures it out, the ambulance has already missed its window. It doesn't scale.
2. The "Trained Robot" (Reinforcement Learning): Imagine teaching a robot driver by showing it millions of hours of traffic videos. Eventually, the robot learns how to drive.
   • The Flaw: The robot is rigid. If you train it on a sunny day with light traffic, it might crash when it sees a rainy day with heavy traffic. It takes a long time to "study" (train), and it can't handle new, weird situations well.

The Solution: The "Smart School of Fish"

The authors propose a system where every car is smart, but only looks at its immediate neighbors. There is no central brain, and no need for years of training.

Think of a school of fish or a flock of birds. They don't have a leader shouting orders. Instead, every bird just follows three simple rules based on the birds right next to it:

1. Don't crash into the neighbor.
2. Don't move too far away from the group's average speed.
3. If an emergency vehicle is coming, gently shift to make room.

How SDVC works in this "School of Fish" style:

• Local Vision: Each car only talks to the cars within a 400-meter range (about 4-5 football fields). It doesn't need to know what's happening 10 miles away.
• The "Influence" Check: Before making a move, a car asks: "Is the ambulance coming? Is the car in front of me slowing down?"
  • If the answer is No, the car just keeps cruising. No changes needed.
  • If the answer is Yes, the car runs a quick mental math check to see the best way to move (speed up, slow down, or change lanes) to let the ambulance pass without causing a ripple effect of panic.
• The "Handshake" (Conflict Resolution): Sometimes, two cars might decide to move into the same empty lane at the same time. In the old days, this would cause a crash. In SDVC, the cars have a quick "handshake." They form a tiny, temporary team (a coalition) to decide who goes first. The ambulance always wins, but the ordinary cars negotiate who moves slightly to avoid a collision.

Why is this a Game-Changer?

1. It's Instant: Because every car does its own small math, there is no waiting for a central computer. It happens in milliseconds—faster than a human can blink.
2. It Scales Forever: Whether there are 20 cars or 2,000 cars, the system works the same way. Each car only cares about its neighbors. Adding more cars doesn't slow down the "brain" because there is no single brain.
3. It's Safe: The system includes a "safety net" that guarantees cars won't crash, even if they are making decisions on the fly.
4. It Adapts: Unlike the "Trained Robot," this system doesn't need to be retrained for a new city or a new type of traffic jam. It figures it out on the spot using logic.

The Results

The researchers tested this using real traffic data from German highways.

• The "Super-Conductor" methods crashed or failed when the traffic got too heavy.
• The "Trained Robot" method caused accidents because it got confused by new traffic patterns.
• The SDVC "School of Fish" let the ambulance zoom through, kept the ordinary cars moving smoothly, and had zero accidents in every single test, even on roads with 5 lanes and heavy traffic.

In short: Instead of trying to control the whole ocean with one giant net, this method teaches every drop of water how to flow around the obstacle naturally, ensuring the ambulance gets through while keeping the ocean calm.

Technical Summary

1. Problem Statement

The rapid transit of Emergency Vehicles (EMVs) is critical for saving lives and minimizing property damage. While EMVs have legal priority, their passage often disrupts ordinary traffic (OVs), leading to inefficiencies and potential safety hazards.

• Current Limitations: Existing control strategies generally fall into two categories:
  1. Centralized Mathematical Solvers: These provide optimal solutions but suffer from high computational costs, making them impractical for large-scale traffic scenarios (e.g., solving for hundreds of vehicles).
  2. Reinforcement Learning (RL): These methods offer distributed control but require extensive pre-training (10–15 hours), struggle to generalize to varying traffic densities and road configurations, and often have limited prediction horizons (e.g., only effective within 100m of the EMV).
• Goal: The paper aims to develop a control method that ensures rapid EMV transit while minimizing the impact on OVs, characterized by low computational cost, high scalability, real-time decision-making, and robustness across diverse traffic conditions without pre-training.

2. Methodology: Scalable Distributed Vehicle Control (SDVC)

The authors propose SDVC, a distributed control framework where vehicles adjust behaviors online using only local information (vehicles within communication range) rather than global traffic data. The method consists of three core components:

A. Influence Judgment

Each Ordinary Vehicle (OV) determines if it needs to adjust its behavior based on two conditions:

1. Safety Violation: Predicted future states of the OV and its neighbors violate safety distance constraints. The prediction horizon ($F_{j,n}$) is adaptive: it is longer when interacting with EMVs (which cannot decelerate) and shorter when interacting with other OVs.
2. Speed Deviation: The OV's speed deviates significantly from the mean speed of its current lane. This ensures that only vehicles causing or facing significant disruption adjust their behavior, minimizing unnecessary changes.

B. Strategy Function

When an OV is "influenced," it evaluates feasible next states (speed, lane, position) using a local strategy function ($F_n$) to select the optimal candidate. The function minimizes:

1. Behavioral Cost: The magnitude of speed changes and lane changes ($f_1$).
2. Coordination Cost: The deviation of the vehicle's speed from the lane's mean speed ($f_2$). Theoretically, minimizing this local deviation approximates minimizing the global coordination cost.
3. Penalty Term: A binary penalty ($f_3$) is applied if a candidate state violates safety constraints or reduces the vehicle's speed below a critical efficiency threshold.

C. Distributed Conflict Resolution

Since vehicles make decisions independently, conflicts (e.g., two vehicles choosing the same lane/space) can occur. SDVC resolves this via Coalition Formation:

1. Initialization: Vehicles identify neighbors with conflicting candidate states and form a "conflict coalition."
2. Priority-Based Resolution: Within a coalition, a central vehicle is selected based on priority (EMVs > OVs with fewer feasible options). The central vehicle computes the optimal states for all coalition members to ensure safety.
3. Iterative Expansion: If conflicts persist, the coalition expands to include the nearest vehicle to increase the solution space, ensuring deterministic safety guarantees without a single point of failure.

3. Key Contributions

1. Scalable Distributed Control: The method operates online using only local information, eliminating the need for global state knowledge or pre-training.
2. Theoretical Equivalence: The authors prove that minimizing the local strategy function is approximately equivalent to minimizing the global optimization objective. This validates that local decisions lead to near-optimal global outcomes.
3. Deterministic Safety: Unlike RL methods which offer probabilistic safety, SDVC provides deterministic safety guarantees through its conflict resolution mechanism, eliminating single points of failure.
4. Zero Training Time: The method requires no training, allowing immediate deployment and natural adaptability to varying traffic densities and road configurations.

4. Experimental Results

Experiments were conducted using real-world traffic data from the HighD dataset (German highways) and compared against state-of-the-art baselines: CRHA (Centralized Mathematical Solver) and GSAC (Graph Convolutional Soft Actor-Critic, an RL method).

• Solution Quality: SDVC achieved significantly lower "impact scores" (fewer lane changes and speed fluctuations for OVs) compared to GSAC (reducing impact by 33–59%) and matched or outperformed CRHA in small-scale scenarios while remaining feasible in large-scale ones.
• Scalability:
  • Large-Scale: In scenarios with hundreds of vehicles (up to 436 OVs) and long routes, CRHA failed to produce results due to computational explosion, while SDVC successfully planned routes.
  • Lane Configurations: SDVC maintained stable performance across 3, 4, and 5-lane roadways. In contrast, GSAC performance degraded significantly in 4 and 5-lane scenarios due to poor generalization.
• Real-Time Performance: SDVC achieved decision latencies of 2.5–3.1 ms, well below the human visual reaction time (200 ms), satisfying real-time requirements.
• Safety: SDVC maintained a 0% collision rate across all tested scenarios (varying densities, speeds, and lane counts). GSAC IDM exhibited collision rates up to 11% in high-density or heterogeneous scenarios.

5. Significance

This work addresses the fundamental trade-off between optimality and scalability in emergency vehicle transit. By proving that local decision-making can approximate global optimality, the paper offers a practical solution for real-world deployment where traffic conditions are dynamic and unpredictable. The method's ability to handle large-scale, multi-lane, and heterogeneous traffic without pre-training makes it a robust candidate for future Intelligent Transportation Systems (ITS), particularly for connected and autonomous vehicle (CAV) environments.