Organizations: Fujitsu Limited, Kawasaki, Japan · Fujitsu Research of America, Pittsburgh, PA, USA · Department of Civil and Environmental Engineering, Carnegie Mellon University, Pittsburgh, PA, USA · Heinz College of Information Systems and Public Policy, Carnegie Mellon University, Pittsburgh, PA, USA
Traffic digital twins, which inform policymakers of effective interventions based on large-scale, high-fidelity computational models calibrated to real-world traffic, hold promise for addressing societal challenges in our rapidly urbanizing world. However, conventional fine-grained traffic simulations are non-differentiable and typically rely on inefficient gradient-free optimization, making calibration for real-world applications computationally infeasible. Here we present a differentiable agent-based traffic simulator that enables ultra-fast model calibration, traffic nowcasting, and control on large-scale networks. We develop several differentiable computing techniques for simulating individual vehicle movements, including stochastic decision-making and inter-agent interactions, while ensuring that entire simulation trajectories remain end-to-end differentiable for efficient gradient-based optimization. On the large-scale Chicago road network, with over 10,000 calibration parameters, our model simulates more than one million vehicles at 173 times real-time speed. This ultra-fast simulation, together with efficient gradient-based optimization, enables us to complete model calibration using the previous 30 minutes of traffic data in 455 s, provide a one-hour-ahead traffic nowcast in 21 s, and solve the resulting traffic control problem in 728 s. This yields a full calibration--nowcast--control loop in under 20 minutes, leaving about 40 minutes of lead time for implementing interventions. Our work thus provides a practical computational basis for realizing traffic digital twins.
Figures & tables
Figure 1: Framework of the differentiable agent-based traffic simulation for traffic nowcasting and control. The framework consists of two main stages: calibration and nowcasting/control. (1) Calibration: The simulation runs with initial conditions X(t=0) , optimizable behavioral parameters ( θb ), and fixed environmental parameters ( θe ). By minimizing the loss ℓ between the simulated traffic state y^(t=T) and ground-truth observations y(t=T) via gradient-based optimization using the gradient ∂ℓ/∂θb , the behavioral parameter is calibrated to θb∗ , accurately replicating observed traffic dynamics. (2) Nowcasting and Control: Using the calibrated parameter θb∗ , the current traffic state X(t=T) , and the current environmental parameter θe , the model performs nowcasting to predict the near-future state y^(t=T+ΔT) . To control traffic, a desired state y∗ is set. The gradient with respect to the environmental parameter, ∂ℓ/∂θe , is then calculated to determine the optimal interventions θe∗ (e.g., dynamic pricing, dynamic speed limit control) that guide the system towards this desired state.
Figure 2: Calibration results on the Chicago Sketch network. (a) Cumulative traffic counts from the mean-parameter simulation and the calibrated simulation, compared against ground-truth counts. (b) Five-minute traffic counts from the mean-parameter simulation and the calibrated simulation, compared against ground-truth counts. (c) Estimated β values compared against ground-truth β values. Because the choice probability of the next links is determined only by differences in utility, β is normalized for comparison. For clarity, we display only the β values of links at junctions connected to the top-10 links by traffic volume.
Figure 3: Traffic nowcasting results on the Chicago Sketch network. (a) Estimated cumulative traffic counts at nowcast horizons of 5, 10, 30, and 60 minutes from the current state. (b) Estimated cumulative traffic counts on links at the 60-minute nowcast horizon compared against ground-truth counts. (c) JIT compilation time for different nowcast horizons. (d) Forward simulation time for different nowcast horizons. In panels c and d, whiskers indicate the standard deviation across five trials. The dashed lines represent the regression results, indicating the linear scalability of the nowcasting computation.
Figure 4: Traffic control results on the Chicago Sketch network. (a) Network view highlighting the target link to be controlled via cost changes. (b) Optimized cost changes designed to halve the cumulative traffic count on the target link. (c) Cumulative traffic counts on the target link with and without control (cost changes). The dashed line represents the target value (50% reduction).
Figure S1: Effect of the trajectory grafting on parameter estimation. (a) Experimental setup. A single agent travels from Link 1 to Link 2 over a total duration of t=10s . The link lengths are L1=L2=15m . The ground-truth trajectory is generated with free-flow speeds u1=1.5m/s on Link 1 and u2=1.0m/s on Link 2, resulting in a single observation of the agent at position x=0m on Link 2 at t=10s . From this observation, we estimate u1 and u2 by minimizing the squared error between the observed and simulated positions. (b) Optimization results without TG (left) and with TG (right). The solid and dashed lines represent estimated and true values, respectively. The results for u1 and u2 are shown in different colors.
Figure S2: Calibration results on the Sioux Falls network. (a) Cumulative traffic counts from the mean-parameter simulation and the calibrated simulation compared against the ground truth. (b) Traffic counts (every 5 minutes) from the mean-parameter simulation and the calibrated simulation compared against the ground truth. (c) Effect of the platoon size Δn on the loss (estimation quality). (d) Effect of the platoon size Δn on the calibration time. The dots and whiskers represent the measured values and their standard deviations over 10 trials, respectively.
Figure S3: Network structures used in this study. (a) Sioux Falls network. (b) Chicago Sketch network. Different node colors represent different inflow and outflow conditions. Bi-directional arrows indicate that there are inbound and outbound links (two links) between two nodes. The numbers of nodes/links for the Sioux Falls and Chicago networks are 48/100 and 993/2571, respectively.
Calibrating microscopic traffic models for digital twins is an expensive black-box optimization problem: tuning car-following and lane-changing parameters requires a full simulation run, affording only a tight budget per recalibration window. Matching raw trajectories yields a rugged objective that sparse surrogates cannot learn, reducing sequential acquisition to near-random probing. We present FLAT, which couples what to optimize with where to sample next. An eight-dimensional behavioral fingerprint smooths the parameter-error landscape, making the objective learnable from a few dozen samples; annealed lower-confidence-bound (LCB) acquisition then spends each remaining run where it most reduces error. The surrogate, interchangeable among a Gaussian process (GP), random forest (RF), or multi-layer-perceptron (MLP) ensemble, plugs into the same LCB loop. Across six heterogeneous real-world scenes, FLAT-GP achieves the lowest scene-averaged behavioral error, winning 6/6 scenes against SPSA, GA, and CMA-ES and 5/6 against TPE under the matched budget. Some baselines need up to 4.4 times more simulations to match. Ablations show objective choice shifts final behavioral error by 81% on average, removing sequential LCB raises the six-scene mean by 20%, and surrogate choice shifts it by at most 4.2%, confirming gains trace to objective geometry and sequential allocation rather than surrogate capacity.
Haopeng Deng, Shuo He, Dayuan Wang
School of Future Transportation Guangzhou Maritime University Guangzhou 510725, China · Department of Civil and Environmental Engineering Imperial College London London SW7 2AZ, United Kingdom · Department of Biostatistics University of Florida Gainesville, FL 32601, USA
Supervised open-loop training has been widely adopted for training traffic simulation models; however, it fails to capture the inherently dynamic, multi-agent interactions common in complex driving scenarios. We introduce RLFTSim, a reinforcement-learning-based fine-tuning framework that enhances scenario realism by aligning simulator rollouts with real-world data distributions and provides a method for distilling goal-conditioned controllability in scenario generation. We instantiate RLFTSim on top of a pre-trained simulation model, design a reward that balances fidelity and controllability, and perform comprehensive experiments on the Waymo Open Motion Dataset. Our results show improvements in realism, achieving state-of-the-art performance. Compared with other heuristic search-based fine-tuning methods, RLFTSim requires significantly fewer samples due to a proposed low-variance and dense reward signal, and it directly addresses the realism alignment issue by design. We also demonstrate the effectiveness of our approach for distilling traffic simulation controllability through goal conditioning. The project page is available at https://ehsan-ami.github.io/rlftsim.
This article outlines a new framework of traffic light optimization through a digital twin of the transport infrastructure, managed by agentic AI to ensure real-time autonomous decisions. The framework relies on physical sensors and edge computing to measure real-time traffic information and simulate traffic flow in a constantly updated digital twin. The traffic light is automatically controlled through the digital twin according to traffic congestion, travel delay and traffic patterns. This approach is implemented as a three-layer system: perception, conceptualization and action. The perception layer receives data on physical systems; the conceptualization layer uses LangChain to process the data; and the action layer links to the Model Context Protocol (MCP) and traffic management APIs to implement optimised traffic signal control algorithms. The results show that the framework minimizes waiting time at traffic lights and positively affects the effectiveness of the entire traffic flow, which is better than the fixed-time and reinforcement learning-based baselines.
Salman Jan, Toqeer Ali Syed, Shahid Kamal +2
Center for Advanced Analytics, CoE for Artificial Intelligence, Faculty of Computing and Informatics, Multimedia University, Cyberjaya, Malaysia · Faculty of Computer Studies, Arab Open University-Bahrain, A’Ali, Bahrain · Faculty of Computer and Information System, Islamic University of Madinah, Madinah, Saudi Arabia +1