Traffic control in the era of automated vehicles
The advance of automated vehicles promises great potential for improving safety and efficiency of traffic flow. Using extensive real-time sensor data, significantly more information about the current traffic state can be gathered and utilized for optimal traffic control. This academic cooperation between TUM and BMW focuses on developing innovative approaches for designing future traffic systems.
Focus Group: Artificial Intelligence in Traffic Engineering and Control
Dr.-Ing. Felix Rempe (BMW, Autonomous Driving), Alumnus Rudolf Diesel Industry Fellow
Athanasia Karalakou (TUM), Doctoral Candidate
Host: Prof. Klaus Bogenberger (TUM)
The Focus Group Artificial Intelligence in Traffic Engineering and Control was founded at the end of 2022 within the Rudolf Diesel Fellowship to connect researchers from the disciplines of automated driving, traffic engineering, and machine learning.
The first research direction focused on the question: How can data from automated vehicles be leveraged to improve traffic efficiency via the provisioning of next-level traffic state information? In state-of-the-art transportation networks, data is collected in basically two ways: via static road installation, e.g., inductive loop sensors, or via probe data, e.g., GNSS data from road users. The first category of sensors is able to perceive the traffic state accurately, but only at one location in the road network. The second category is, as of now, able to provide data for the entire road network, but with limited accuracy and completeness. With decreasing connectivity costs and advances in sensor systems, more and more vehicle sensor data is available. This type of data, in combination with novel algorithms for processing, can be expected to enable the provisioning of next-level scalable and accurate traffic information for various types of applications in traffic engineering, ranging from long-term traffic planning over real-time control to optimal lane-level routing, ultimately helping to reduce congestion and its costs.
However, the sparsity of probes remains a challenge since the penetration rate of submitting vehicles, even in the best scenarios, will be relatively low due to the heterogeneity of fleets and distributed ownership. Furthermore, traffic conditions, especially in urban road networks, change dynamically. Using empirical sensor data that stems from a large fleet of BMWs equipped with cameras and radars as well as data from vehicle simulations, we developed novel approaches for predicting traffic information at locations and points in times where no measurements are available. First, we developed novel variants of a neural network that can estimate spatio-temporal traffic conditions in road networks based on a few sensing vehicles [1] Summarized, the results show that this extended data, processed with current machine learning approaches, already allow insights into traffic conditions in a level of detail that has not been possible before: Traffic density, flow, and speed are available on a lane level compared to road-level-fine traffic speeds as of now. Second, we contributed to research on so-called network fundamental diagrams (NFDs), which link traffic density, speed, and flow in homogeneous subgraphs of traffic networks enabling traffic control at city scale. In our research, we demonstrated how empirical data from vehicle fleets, as well as data from fully automated vehicles, can be utilized for the generation of NFDs [2] [3].
While the first research direction focused on traffic state estimation on a macroscopic level, the second research direction targeted optimal traffic control on the microscopic level. With the increasing automation of vehicles, there is the potential to rethink our road infrastructure, which was primarily built to ensure a safe drive for humans in interaction. Via vehicle-to-everything (V2X) communication, together with the reduction of potentially obsolete safety buffers, traffic flow might become significantly more efficient.
Figure 1
In the concept of so-called lane-free traffic, the core idea is to relax current static lane designs since lanes are often much wider than necessary or underutilized, wasting precious space. In our Focus Group, we developed strategies to pave the road for the realization of lane-free traffic. For instance, we experimented with a dynamical adaptation of the lane width in freeway weaving sections, which is a common nucleus for traffic breakdown and subsequent congestion. It turned out that adding or removing lanes can reduce on-demand
conflicts between road users, and thus the capacity of the road can be increased. Another scenario that we looked into is lane-free roundabouts. This type of road infrastructure element is a common building block in road networks. While an optimal behavior of automated vehicles in intersections is studied by many other researchers, e.g., by the group of Prof. Papageorgiou (TUM ambassador since 2021), traffic flow in roundabouts may likewise benefit from lane-free concepts. Results of our research on roundabouts are (for now as of 2025), twofold: an evaluation of microscopic models for vehicle agents in simulations to be used in lane-free roundabouts; and a novel methodology to control vehicles in lane-free conditions based on deep reinforcement learning [4] [5] .
In summary, thanks to support from the TUM IAS, novel approaches for cooperative traffic state estimation and traffic control have been developed within this research group. Several publications on international conferences and journals were completed. Finally, a workshop with international speakers and more than 50 participants took place at the end of 2025, concluding the activities and inspiring new ideas. Among the German and international guests, we hosted Prof Papageoriou, the founder of the lane-free traffic concept, and Dr. Martin Treiber, two of the leading experts worldwide in traffic control, as well as Prof. Aleksandar Stevanovic from the University of Pittsburgh. The results impact discussions on policy making for optimal traffic control, as well as the development of safe driving strategies at BMW.
In close collaboration with Yunfei Zhang, M. Sc. (TUM)
Figure 2
[1]
A. Karalakou, F. Rempe, L. Kessler, and K. Bogenberger (2024).
[2]
Y. Zhang, M. Ilic, and K. Bogenberger (2023).
[3]
Y. Zhang, F. Rempe, F. Dandl, G. Tilg, M. Kraus, and K. Bogenberger (2023).
[4]
A. Karalakou, M. Rostami-Shahrbabaki, F. Rempe, and K. Bogenberger (2025)
[5]
A. Karalakou et al. (2024)
Selected publications
- A. Karalakou, F. Rempe, L. Kessler, and K. Bogenberger, “Lane-based traffic state estimation on freeways using empirical automated vehicle data,” in Proc. 103rd Annu. Meeting Transp. Res. Board (TRB), Washington, DC, USA, 2024.
- Y. Zhang, M. Ilic, and K. Bogenberger, “A novel concept of traffic data collection and utilization: Autonomous vehicles as a sensor,” in 2023 IEEE 26th Int. Conf. Intell. Transp. Syst. (ITSC), Bilbao, Spain, 2023, pp. 3887–3892.
- Y. Zhang, F. Rempe, F. Dandl, G. Tilg, M. Kraus, and K. Bogenberger, “Network fundamental diagram based dynamic routing in a clustered network,” in 2023 8th Int. Conf. Models Technol. Intell. Transp. Syst. (MT-ITS), Nice, France, 2023, pp. 1–7.
- A. Karalakou, M. Rostami-Shahrbabaki, F. Rempe, and K. Bogenberger, “A deep reinforcement learning approach for controlling autonomous vehicles in lane-free roundabouts,” in 2025 IEEE Intell. Veh. Symp. (IV), Cluj-Napoca, Romania, 2025, pp. 2348–2354.
- A. Karalakou et al., “Evaluation and validation of microscopic mdels for CAVs in lane-free roundabouts,” in 5th Symp. Manage. Future Motorway Urban Traffic Syst. (MFTS), Heraklion, Greece, 2024.