Data-driven generalized perimeter control: Zürich case study

Alessio Rimoldi; Carlo Cenedese; Alberto Padoan; Florian Dörfler; John Lygeros

eess.SYMar 17, 2026

Data-driven generalized perimeter control: Zürich case study

Authors: Alessio Rimoldi, Carlo Cenedese, Alberto Padoan, Florian Dörfler, John Lygeros

Organizations: Automatic Control Laboratory, ETH Zürich · Delft Center for Systems and Control, TU Delft · Dep. of Electrical and Computer Engineering, UBC

Abstract

Urban traffic congestion is a key challenge for the development of modern cities, requiring advanced control techniques to optimize existing infrastructures usage. Despite the extensive availability of data, modeling such complex systems remains an expensive and time consuming step when designing model-based control approaches. On the other hand, machine learning approaches require simulations to bootstrap models, or are unable to deal with the sparse nature of traffic data and enforce hard constraints. We propose a novel formulation of traffic dynamics based on behavioral systems theory and apply data-enabled predictive control to steer traffic dynamics via dynamic traffic light control. A high-fidelity simulation of the city of Zürich, the largest closed-loop microscopic simulation of urban traffic in the literature to the best of our knowledge, is used to validate the performance of the proposed method in terms of total travel time and CO2 emissions.

Figures & tables

Figure 1: The city is divided into $p$ regions homogeneous in average traffic density (the hexagons) and the demand $\boldsymbol{d}$ among them. A detail of sensors $\mathbf{s}_{j}$ (blue diamonds) and traffic lights $\mathbf{l}_{j}$ (green circles) locations within region $j\in\mathbf{p}$ is shown.

Panel of Figure 2 — Figure 2: Example of a four-legged intersection and compatibility of traffic flows. On the left the structural components of the intersection are highlighted for clarity. On the right the difference between compatible and antagonistic flows can be observed.

Figure 3: Example of a duty cycle comprising six phases divided in active and passive. Active phases (slowromancapi@,slowromancapiv@) are grouped in the leftmost column and highlighted in green, while passive phases (slowromancapii@,slowromancapiii@,slowromancapv@,slowromancapvi@) can be found in the two rightmost columns highlighted in red. Active phases have at least one ’g’ state in the tuple.

Figure 4: Different control input values

\lambda_{\ell}

result in different active and passive phase durations, that is

\delta_{a}

and

\delta_{p}

, for a traffic light

\ell\in\mathbf{l}

. The bottom graph shows a possible input trajectory for the

\ell

-th traffic light spanning three duty cycles, the top graph shows proportion of duty cycle time assigned to the active and passive phases as a result of the input.

Figure 5: Commutative diagram of the relationships between

\lambda

\rho

and

\phi

. The relationship between

\rho

and

\phi

is assumed to be a MFD similar to Figure 6 .

Figure 6: The MFD for the region of Wiedikon, Zürich (solid green line) simulated below. The critical density

\rho_{\textup{cr},i}

is the density where the flow is maximum, and the maximal density of region

\rho_{\textup{max},i}

is the density where the flow becomes zero and region

i

gridlocks.

Panel of Figure 7 — Figure 7: Left panel shows the demand profile, the vehicles move from the outer to the inner region. The network is partitioned into the outer region (solid magenta line) and in inner region (dashed green line). Black dots represent the controlled traffic lights.

Panel of Figure 8 — Figure 8: Evolution of the density (left panel) and flow (right panel) in inner region of the lattice network.

	Baseline	MPC	DeePC	% variation
Travel Time (min)	27.5	26.03	$\boldsymbol{24.93}$	- 9.34%
Waiting Time (min)	25.33	22.24	$\boldsymbol{21.39}$	- 15.55%
CO Emissions(g)	259.65	239.59	$\boldsymbol{229.5}$	- 10.92%
CO2 Emissions (g)	4,351.16	4,084.12	$\boldsymbol{3,917.79}$	- 9.95%
HC Emissions (g)	1.28	1.19	$\boldsymbol{1.14}$	- 10.37%
PMx Emissions (g)	0.11	$\boldsymbol{0.1}$	$\boldsymbol{0.1}$	- 8.95%

Table 1: Average metrics for the lattice network

Panel of Figure 9 — Figure 9: Left panel shows the Zürich urban traffic network embedded in SUMO . Actuated intersections are highlighted in red. In the corner, one of the controlled intersections, Bellevue. Right panel shows the region partitioning found using the snake clustering algorithm, the Wiedikon region is highlighted in red.

Figure 10: The demand profile directed to the Wiedikon region broken down by region of origin. The color coding used matches the one in Figure 9(b) .

Panel of Figure 11 — Figure 11: Evolution of density (left panel) and flow (right panel) under control of DeePC and MPC for the Wiedikon region. The critical density value $\rho_{\textup{cr}}$ (red dashed line) is reported in the left panel, the line represents the threshold for congestion of the system.

Panel of Figure 12 — Figure 12: Evolution of density (left panel) and flow (right panel) under control of DeePC with different periods for the Wiedikon region.

	Baseline	DeePC 6 $\Delta_{\textrm{dc}}$	DeePC 3 $\Delta_{\textrm{dc}}$	MPC	DeePC 1 $\Delta_{\textrm{dc}}$
Travel Time (min)	45.23	39.07	37.94	38.51	$\boldsymbol{37.05}$	-18.08%
Waiting Time (min)	28.27	23.31	22.62	23.07	$\boldsymbol{21.87}$	-22.60%
CO Emissions(g)	16.91	15.84	15.41	16.14	$\boldsymbol{15.67}$	-7.33%
CO2 Emissions (g)	4,992.15	4,379.17	4,253.91	4,338.89	$\boldsymbol{4,184.46}$	-16.17%
HC Emissions (g)	0.11	0.11	0.10	0.11	0.11	-0.00%
PMx Emissions (g)	0.22	0.21	0.21	0.21	0.21	-4,54%

Table 2: Average Metrics Zürich

Component	PC1	PC2	PC3	PC4	PC5	PC6	PC7	PC8	PC9	PC10
EVR	$0.477$	$0.103$	$0.051$	$0.037$	$0.033$	$0.026$	$0.022$	$0.021$	$0.020$	$0.018$

Table 3: Explained variance ratios of the components, each value represents the percentage of variation in the control input that can be explained by the corresponding principal component.

Panel of Figure 13 — Figure 13: Left panel shows the evolution of the first three PCA components. Right panel shows the loadings value associated to PC1.

Figure 14: Positions of the highly influential actuators. Left panel shows the positions of the ten highest loading value (positively correlated) actuators highlighted in red. Right panel shows the positions of the ten lowest loading values (inversely correlated) actuators highlighted in magenta.

Panel of Figure 15 — Figure 15: Input time series of the two most influential actuators respectively positively (left panel) and negatively correlated actuators.

Figure 16: Comparison between the density-flow data distribution of Wiedikon obtained under the static control policy (gray dots) and under the DeePC control policy (blue squares). The MFD s and critical densities

\rho_{\textup{cr}}

are shown in the corresponding colors for both distributions.

Data-Driven Control

Data-driven control uses measured data and learned models or policies to control systems when accurate first-principles models are unavailable.

See all 36 papers

TrafficClaw: A Generalizable LLM Agent in the Unified Physical Environment for Urban Traffic Control
Apr 19, 2026Siqi Lai, Pan Zhang, Yuping Zhou +3Agentic RL Language Model-Based Control
The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, Guangdong, China · Shandong University, Jinan, Shandong, China
HiLLTS: Zero-Shot Hierarchical LLM-Guided Traffic Signal Control for Sustainable Transportation
Jul 17, 2026Yue Ding, Tendai Mukande, Mingming LiuLanguage Model-Based Control Intelligent Transportation Systems
Research Ireland Centre for Researching Training in Machine Learning at Dublin City University
Momentum Based Reward Design for Low Emission Traffic Signal Control
May 28, 2026Chinmay Mundane, Amith Manoharan, Arun Kumar SinghReinforcement Learning Reward Shaping
Institute of Technology, University of Tartu

Apr 19, 2026J/K move · Enter open · S save

TrafficClaw: A Generalizable LLM Agent in the Unified Physical Environment for Urban Traffic Control

Siqi Lai, Pan Zhang, Yuping Zhou, Jindong Han+2

The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, Guangdong, China · Shandong University, Jinan, Shandong, China

Agentic RL Language Model-Based Control Intelligent Transportation Systems Traffic Signal Control

Open paper AskPDF arXiv

Large language model (LLM) agents have shown strong capabilities in long-horizon reasoning, tool use, and decision-making in digital environments, yet extending them to physically grounded systems remains challenging. Unlike web, code, or game environments, where objectives are often weakly coupled, physical systems evolve through tightly coupled dynamics in which local interventions propagate across interacting subsystems over time. Urban traffic control exemplifies this challenge, as traffic signals, freeways, public transit, and taxi systems continuously interact through shared spatial infrastructure and temporal mobility demand. Existing optimization, reinforcement learning (RL), and LLM-based approaches are largely designed for isolated subsystems, limiting coordinated reasoning and system-level optimization. We propose TrafficClaw, a LLM-based generalizable traffic control agent for physical urban systems. TrafficClaw operates within a unified traffic environment that exposes coupled urban dynamics and feedback, performs executable spatiotemporal reasoning with persistent memory for long-horizon adaptation, and leverages multi-stage agentic RL for coordinated system-level optimization. Experiments across three metropolitan regions and six traffic-control tasks demonstrate strong generalization, robustness, and cross-subsystem coordination. Our project is available at https://github.com/usail-hkust/TrafficClaw.

Abstract

Data-Driven Control

Explore similar work

TrafficClaw: A Generalizable LLM Agent in the Unified Physical Environment for Urban Traffic Control

HiLLTS: Zero-Shot Hierarchical LLM-Guided Traffic Signal Control for Sustainable Transportation

Momentum Based Reward Design for Low Emission Traffic Signal Control