Practical Training "Planning Robust Behavior Generation for Autonomous Vehicles"

Organizer: Klemens Esterle, Patrick Hart

Contact: esterle (at)

Modul: IN2106, IN4251

Registration: Via Matching System and Submitted Application

Type: Practical Training

Semester: Summer Semester 2020

News (Registration)

Registration is only possible by submitting an application of interest over email and respectively applying through matching system.


We are a group of Phd students working working on behavior planning. For that, we work both in simulation and on our own prototype vehicle.


In this practical course, you will work on one of the remaining key challenges in autonomous driving: Robust behavior generation in the face of behavioral uncertainty. Given routing information, a static map and the motion history of all other agents, behavior planning deals with the problem of finding a continuous, collision-free and dynamically feasible time-dependent motion while considering traffic regulations, social conventions and time constraints. Scenarios with high interactions between many participants, such as merging in dense traffic, require the negotiation with other participants. To achieve robust behavior, the unknown intentions of other participants need to be reliably estimated and incorporated into the planning process. Handling such behavioral uncertainty is computationally demanding due to an exponentially increasing set of possible maneuvering options. Though several approaches have been proposed in the past, no method has demonstrated all necessary requirements for autonomous driving at SAE level 3 and above. In this practical course, we develop and implement, in teams, state-of-the-art behavior generation algorithms for autonomous vehicles. We select methods from different fields, such as Deep Reinforcement Learning, Search-Based Methods and Formal Methods. In a final contest, we will compare the different developed algorithms and draw conclusions.

The implementation is based on our open source simulation platform providing visualization and data handling. Thus, students can fully concentrate on designing and improving their behavior generation module.


Weekly Meetings: Weekly meetings during the semester. These weekly meetings are mandatory for all participating students.

Location: The practical course will take place at fortiss (Research Institute of the Free State of Bavaria associated with Technical University of Munich), located at metro station U6 Nordfriedhof.

Expected previous knowledge: Students should have significant experience with Python and C++.