Despite the tremendous progress across nearly all vision tasks, the utopia of autonomous cars that navigate solely using vision-based sensor modalities such as camera or lidar has not yet materialized. Even though the progress in perception has led to significant advances in fundamental prerequisites such as robust detection of surrounding vehicles or pedestrians/cyclists, the rise of autonomous vehicles is held back by the conflict between high safety requirements and an infinite diversity of potential traffic scenarios. Throughout this dissertation, we focus on challenging off-highway environments, i.e., urban traffic characterized by complex interactions among self-driving vehicles and surrounding traffic, as well as vulnerable road users. Diverse road topologies, unsignaled intersections, occlusions caused by buildings and parked vehicles, or construction zones make this particularly challenging. To efficiently navigate crowded urban environments, self-driving vehicles must understand their environment beyond accurately perceiving objects. They must focus on objects relevant to the driving task and forecast their surroundings’ behavior. The result of this so-called prediction task can then be leveraged to plan a safe and efficient trajectory. This dissertation addresses prevalent problems in the two interconnected fields of prediction and planning. We analyze how seeing them as tightly coupled tasks is paramount to reason about interactive maneuvers and demonstrate how the robustness of vehicle trajectory prediction can be improved. In addition to increased data efficiency, our methods yield more reproducible and realistic distributions over the future behavior of traffic agents. Then, we show how prediction methods can be repurposed as strong learning-based planning baselines. We also introduce a novel rule-based planning algorithm that achieves state-of-the-art performance in realistic urban traffic scenarios. Finally, we put a comprehensive set of rule-based and learning-based methods to the test in a novel benchmark centered around highly interactive maneuvers and realistic long-tail scenarios. Our results reveal that despite achieving excellent results in regular lane-following scenarios, many methods fail to generalize to these critical cases. Based on our findings, we outline promising avenues for future research that are crucial to enabling real-world autonomy.


    Access

    Download

    Check availability in my library


    Export, share and cite



    Title :

    Data-driven Behavior and Motion Planning for Autonomous Driving in Interactive Urban Environments


    Contributors:

    Publication date :

    2025



    Type of media :

    Miscellaneous


    Type of material :

    Electronic Resource


    Language :

    English




    Path Planning for Autonomous Bus Driving in Urban Environments

    Oliveira, Rui / Lima, Pedro F. / Pereira, Gonçalo Collares et al. | ArXiv | 2019

    Free access


    PATH PLANNING IN AUTONOMOUS DRIVING ENVIRONMENTS

    GYLLENHAMMAR MAGNUS / SIVENCRONA HÅKAN | European Patent Office | 2024

    Free access

    PATH PLANNING IN AUTONOMOUS DRIVING ENVIRONMENTS

    GYLLENHAMMAR MAGNUS / SIVENCRONA HÅKAN | European Patent Office | 2022

    Free access

    Traffic Mirror-Aware POMDP Behavior Planning for Autonomous Urban Driving

    Zhang, Chi / Steinhauser, Florian / Hinz, Gereon et al. | IEEE | 2022