Core Concepts
Reinforcement learning improves safety, efficiency, and stability in mixed traffic.
Abstract
The article discusses the challenges of congestion in traffic due to human-driven vehicles and how Robot Vehicles (RVs) can mitigate these issues. It introduces EnduRL, a framework that incorporates real-world driving profiles into simulations to train RVs using reinforcement learning. The study evaluates safety, efficiency, and stability in two mixed traffic environments (Ring and Bottleneck) at various penetration rates. Results show significant improvements in safety by up to 66%, efficiency by up to 54%, and stability by up to 97% under real-world perturbations.
I. INTRODUCTION
Human-driven vehicles amplify traffic perturbations.
Robot Vehicles (RVs) can mitigate congestion.
EnduRL uses real-world driving profiles for RV training.
II. RELATED WORK
Various control strategies for mixed traffic have been proposed.
Model-based, heuristic-based, and learning-based RVs are evaluated.
EnduRL leverages reinforcement learning for improved performance.
III. METHODOLOGY
Intelligent Driver Model (IDM) assumptions explained.
Sampling real-world perturbations for simulation accuracy.
Reinforcement Learning with Congestion Stage Classifier detailed.
IV. EXPERIMENTS
Evaluation metrics include TTC, DRAC, FE, throughput, CAV, WAR.
Results show improvements in safety, efficiency, and stability.
Comparison with existing RV controllers presented.
V. CONCLUSION AND DISCUSSION
EnduRL enhances safety, efficiency, and stability in mixed traffic.
Practical solution without costly infrastructure upgrades.
Future directions include lane-changing dynamics and vehicle types.
Stats
"Results show that under real-world perturbations...efficiency by up to 54%, and stability by up to 97%."
"Our results show that in Ring...our RV improves both safety and efficiency by up to 54%, and stability by up to 97%."
"Our evaluation metrics include the following...wave attenuation."
"Compared to RL + L...For Ours (20%), multiple follower RVs closely track the leader RV."
Quotes
"In contrast, RVs attenuate the standard perturbation of the leader HV..."
"No other studies impose no artificial approximation or bounds..."
"Our approach offers a practical solution because it solely relies on observations of individual RVs."