Emergency Vehicle-Aware Automated Vehicle Lane Change Decision Methodology Estimated to Provide Approximately 26 Percent Points More Blockage-Free Travel Compared to Rule-Based Lane Change Approaches.
Simulation Study Tested AI-Based Emergency Vehicle-Aware Lane Changing Model Incorporated into Fully Automated Vehicles.
Emergency Vehicle Aware Lane Change Decision Model for Autonomous Vehicles Using Deep Reinforcement Learning
Summary Information
Fully Automated Vehicles (AVs) aim to offer safe and comfortable car travel by minimizing the burden of driving. AVs share the road with various types of vehicles, including Emergency Vehicles (EMVs). When approached by an active EMV, the decision-making process of an AV includes the responsibility of changing lanes, typically handled by the Lane Change Decision (LCD) model to some degree of efficiency. This study proposed an Emergency Vehicle Aware LCD (EMV-LCD) model via utilizing Deep Reinforcement Learning (DRL). The proposed solution, EMV-LCD, was evaluated against a rule-based LCD, in terms of safety and level of cooperativeness with the EMV, using simulation.
METHODOLOGY
This study applies DRL to train an automated vehicle to respond appropriately when approached by an EMV. The decision-making process includes definitions for the vehicle's observations (state), possible actions, and a reward system that guides learning. At each time step, the AV observes its surroundings and selects an action based on the current state. The simulation model consists of a three-lane highway with one EMV, one AV, and several human-driven vehicles. Two types of training episodes are used:
- Episode 1 (Eps-1): Designed to teach the AV to yield when followed by an EMV.
- Episode 2 (Eps-2): Aimed at encouraging the AV to keep the lane ahead clear and avoid unnecessary lane changes when the EMV is detected.
FINDINGS
- Results with the proposed EMV-LCD model showed approximately 26 percent points more blockage-free travel (100 percent and 74 percent) compared to the rule-based lane change approach, considering the average of three desired speed levels.
- The EMV-LCD was also found to take 70 percent less time (14 seconds compared to 46 seconds) to yield the lane when approached by an EMV.
- On average, the proposed model was found to be 1.8 percentage points more collision-free than the rule-based lane change approach.
