Deep Reinforcement Learning for Autonomous Vehicles in Mixed Traffic: Capacity and Fuel Efficiency Gains on NGSIM Data

Summary

This arXiv paper (2603.25328) by Pankaj Kumar, Pranamesh Chakraborty, and Subrahmanya Swamy Peruru studies the control of autonomous vehicles (AVs) in mixed traffic, where multiple objectives—safety, efficiency, comfort, fuel economy, and traffic rule compliance—must be balanced while accounting for heterogeneous human driver behavior. The authors train a Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning agent on the NGSIM highway dataset, enabling realistic interaction with human-driven vehicles. Evaluation across the spectrum from fully human-driven to fully RL-controlled traffic shows that full RL control increases road capacity by approximately 7.52%. Regarding fuel consumption, RL-based AVs improve average fuel efficiency by about 28.98% at higher speeds (above 50 km/h) and 1.86% at lower speeds (below 50 km/h) compared with the Intelligent Driver Model (IDM). The work demonstrates that deep reinforcement learning can meaningfully improve throughput and energy efficiency in mixed autonomy environments using data grounded in real-world highway driving.