Zhang, Yang (Tsinghua University), Bian, Ziyu (Tsinghua University), Liu, Qiyuan (Tsinghua University), Li, ZhiHeng (Tsinghua University)

InverseLight: Inverse Dynamics Enabling Policy Transfer with Limited Data for Traffic Signal Control

Scheduled for presentation during the Video Session "On-Demand Video Presentations" (VP-VP), Saturday, November 22, 2025, 08:00−18:00, On-Demand Platform

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on April 2, 2026

Abstract

Traffic signal control (TSC) plays a vital role in ensuring efficient and safe transportation. While reinforcement learning (RL) has shown promising performance in simulators for TSC, transferring these policies to real-world scenarios remains a major challenge due to the dynamics gap and the high cost of collecting large-scale real-world data. In this paper, we propose InverseLight, a two-stage policy transfer framework that enables effective sim-to-real adaptation with limited real-world samples. Our method leverages an inverse dynamics model trained on a small amount of real-world data to reshape the reward function, allowing a pre-trained RL policy to adapt without modifying its architecture. InverseLight is compatible with existing RL-based TSC methods and can be applied as a plug-and-play module. Experimental results on multiple traffic datasets demonstrate that our approach significantly reduces performance degradation during policy transfer, improving traffic efficiency under diverse conditions.