Balancing safety and efficiency in automated vehicles with risk-based planning
Leon Tolksdorf defended his PhD thesis at the Department of Mechanical Engineering on February 11.
Automated vehicles (AVs) are rapidly becoming part of everyday mobility, yet their widespread deployment brings new challenges. For many years to come, AVs will need to share the road with human drivers, cyclists, and pedestrians - often with limited communication and incomplete knowledge of their surroundings. Sensors may provide noisy data, the intentions of other road users remain uncertain, and traffic situations evolve unpredictably. As a result, AVs have to make critical decisions under uncertainty, aiming to avoid unsafe behaviour caused by either overconfidence or excessive caution. Although safety standards emphasize the need to address uncertainty, the field still lacks agreement on how uncertainty should be integrated into core motion planning algorithms. In his PhD research, Leon Tolksdorf investigates this problem by introducing risk鈥慳ware methods for motion planning and decision鈥憁aking in AVs.
An AV motion planning system must evaluate the likelihood of unsafe events, understand how severe such events could be, and make decisions that balance safety with travel efficiency. To support this, Leon Tolksdorf introduces a risk metric based on the expected severity of a collision, combining both collision probability and collision severity within a single framework.Using this metric, the work explores how stochastic model predictive control (SMPC) can generate AV behavior that adapts to different levels of uncertainty. The findings show that SMPC鈥慴ased planners can effectively balance risk while maintaining efficient progress toward a destination.
Real鈥�time estimation of collision probability
To enable practical use of the proposed risk metric, Tolksdorf presents a computationally efficient algorithm for estimating the probability of collision (POC). The method uses multi鈥慶ircular shape approximations for both the AV and surrounding road users, enabling fast, real鈥憈ime computation, precise POC estimates that never underestimate risk, and significantly faster performance compared to other methods. When used within an SMPC motion planner, the algorithm produces smooth and reproducible trajectories, even when the environment becomes more uncertain.
Extending estimation to collision severity
Beyond estimating collision probability alone, the research expands the method to include collision severity, recognizing that different collision types lead to different levels of harm. This extension makes it possible to evaluate full risk across many collision configurations, while remaining compatible with optimization鈥慴ased motion planning strategies. The result is a flexible and application鈥憆eady approach to representing risk in traffic scenarios.
Evaluating egoistic, altruistic, and collective AV behavior
Risk in traffic is not experienced equally by all road users. To address this, Tolksdorf introduces a perspective鈥慴ased extension of the risk metric that allows an AV to minimize its own risk (egoistic perspective), the risk experienced by others (altruistic perspective), and a balance of both (collective perspective). Simulations across a large database of real and artificial traffic scenarios show that collective behavior achieves the most favorable balance. It supports safe interactions among all road users while preserving travel efficiency, suggesting a promising route toward ethical, socially aware automated driving.
Risk鈥�aware decisions in emergency maneuvers
The proposed risk metric is also applied to emergency responses in the research, specifically choosing between emergency braking and evasive steering when a pedestrian is detected late. Simulation results indicate that risk鈥慳ware decision鈥憁aking more reliably selects the safer maneuver than deterministic approaches. Experimental testing with a pedestrian dummy emerging from a sight obstruction confirms the real鈥憌orld applicability of the method.
Ethical and uncertainty鈥�aware automated driving
Overall, this research demonstrates that risk, and thus uncertainty, can be systematically incorporated in AV algorithms for motion planning and decision-making. The proposed algorithms are computationally efficient for real-time optimization-based motion planning, navigating a fully automated vehicle smoothly through road traffic. By showing how AVs can balance their own travel efficiency with the well鈥慴eing of other road users, the research provides a foundation for advancing ethical, reliable, and intelligent automated mobility.
Title of PhD thesis: . Supervisors: Prof. Nathan van de Wouw, Prof. Christian Birkner and Dr. Arturo Tejada Ruiz.