Physical AI for Traffic Jam Prediction and Control

Motivation

Road traffic congestion remains a persistent challenge in urban and suburban areas worldwide, significantly impacting productivity, fuel consumption, and quality of life. While sudden traffic jams can rapidly emerge even without obvious infrastructure bottlenecks, existing traffic control methods often rely on expensive infrastructure upgrades and reactive mechanisms. This project aims to address congestion proactively by combining an in-depth understanding of sudden jam emergence with reinforcement learning-based, infrastructure-free traffic control solutions.

Contributions

Approach 1: Characterizing Sudden Traffic Jams

• We introduced and formally defined "sudden traffic jams," identifying their emergence from short, intense traffic bursts using empirical loop-detector and Uber Movement speed data across New York City, Nairobi, and São Paulo.

• We derived methods to construct "traffic curves" from coarse-grained data, providing insights into traffic jam dynamics and road capacity utilization, especially under resource-constrained scenarios.

Approach 2: Reinforcement Learning for Congestion Control

• We formulated freeway congestion management as a centralized reinforcement learning problem, using a physics-informed approach to dynamically modulate vehicle speeds and maintain smooth traffic flows.

• We evaluated our RL-based control strategy using high-fidelity simulations on real-world road networks, achieving significant improvements in throughput, flow smoothness, and resilience compared to traditional signaling methods.

These approaches combine seamlessly into a unified pipeline, where jam characterization provides critical insights informing effective RL-driven congestion control strategies.

Results

• Our analysis uncovered that Nairobi experiences significantly longer jam durations per road segment than New York City and São Paulo, indicating critical gaps in existing management approaches and underutilized road capacities.

• The reinforcement learning framework improved network throughput by 5%, reduced average vehicle delay by 13%, and decreased total stops by up to 5%, highlighting substantial gains over baseline methods without requiring infrastructure changes.

Papers

• A. Bhardwaj, R. Asim, S. Chauhan, Y. Zaki, L. Subramanian. Self-Regulating Cars: Automating Traffic Control in Free Flow Road Networks. AAAI Conference on Artificial Intelligence, AI for Social Impact, 2026.  PDF. Code.

• A. Bhardwaj*, S. Iyer*, S. Ramesh, J. White, L. Subramanian. Understanding Sudden Traffic Jams: From Emergence to Impact. Development Engineering: Journal of Engineering in Economic Development, 2023.  PDF.

Collaborators

• Rohail Asim (NYU)

• Sachin Chauhan (IITD)

• Shiva Iyer (NYU / Toyota ITC)

• Jerome White (World Bank)

• Yasir Zaki (NYU)

• Lakshminarayanan Subramanian (Advisor, NYU)