RL for Emergency Vehicle Traffic Optimization

Problem

When an ambulance is stuck in traffic, the seconds lost to surrounding vehicles reacting late can cost lives. Lane-level dynamics are the bottleneck: which vehicle should yield, into which lane, and at what moment. Fixed traffic rules handle this poorly, so we set out to learn the behavior instead.

Approach

We modeled lane selection as a reinforcement learning problem with vehicle-to-vehicle (V2V) communication in the loop. Vehicles share state, which allows the policy to coordinate yielding behavior ahead of the emergency vehicle rather than reacting to it. Policies were trained with Stable-Baselines3 and evaluated in SUMO across multi-modal traffic environments, ranging from grid-world abstractions to road networks inspired by real layouts.

Architecture & Tech

A SUMO simulation stack drives the environments, and a Gym style wrapper exposes lane-level state, including V2V shared observations, to the Stable-Baselines3 agents. The evaluation suite compares the learned policy against three heuristic baselines: SUMO's default lane changing, a fixed lane policy, and a utility weighted best lane heuristic. An empty road run serves as the theoretical lower bound on traversal time.

Results & Impact

• The learned policies cut emergency vehicle traversal time by up to 10.9% against the strongest of the three baselines, on a 2 km four lane network, evaluated over 100 runs per setting.
• I validated the policy against New York City traffic speed data and wrote the peer review response defending the method.
• The work was published as "Improving Lane Level Dynamics for EV Traversal: A Reinforcement Learning Approach" at VEHITS 2024, where it was nominated for the Best Student Paper Award.

What I'd do next

I would stress the policies under partial V2V adoption, where only some vehicles communicate, add sensor noise to the observations, and test sim-to-real transfer on scaled testbeds.