Motion Planning in Dynamic Environments with Application to Self-Driving Vehicles
- Doctoral Thesis
Rights / licenseIn Copyright - Non-Commercial Use Permitted
This thesis is concerned with the development of trajectory planning approaches targeting autonomous driving applications in dynamic environments shared with other traffic participants. The goal is to enable mobile robots to operate in challenging environments, characterized by narrow spaces and close proximity of other agents. With their broad range of private and commercial applications reaching from logistics to valet parking to name a few, driverless vehicles have gained increasing attention in the past years. The complexity of automated driving, however, differs widely with respect to the application domain. In contrast to automated highway driving where the set of admissible actions and different traffic participants is usually small, driverless vehicles in urban environments are faced with a broad variety of semantic information and action choices. In addition, the challenging system kinematics and rotation-variant shape of cars call for system-compliant motion planning algorithms and accurate collision detection with both static elements and other traffic participants in the world. Most state-of-the-art approaches divide the motion planning problem in dynamic environments into prediction and planning. First, the future evolvement of the world is predicted, and in the following the autonomous agent seeks a plan to avoid predicted obstacles at any time. The fact that other traffic participants will adapt their future actions depending on the presence of the robot is neglected in order to make the challenging problem of real-time motion planning in dynamic environments tractable. This simplification, however, gives rise to over-conservative behavior of the autonomous agent, as the latter is rendered responsible for collision avoidance alone. Especially in more densely populated environments with few or weak traffic rules, neglecting the possibility of interaction between the autonomous agent and other traffic participants may lead to reduced navigation performance. This thesis addresses two aspects of motion planning for autonomous vehicles. First, the thesis contributes a non-cooperative, sampling-based motion planning approach, suitable for autonomous vehicle navigation in narrow environments such as parking lots in valet parking applications. The second part is dedicated to cooperative motion planning to improve upon navigation in the presence of other agents. The sampling-based motion planning approach, developed in the first part of the thesis, focuses on planning highly system-compliant motions in order to minimize the divergence between the motion planning and execution layer, and fast and accurate time-aware collision detection to test trajectories of objects of arbitrary shape for collision. Both system-compliant motion planning as well as accurate collision detection is mandatory for narrow environments in order to ensure safety, and avoid the loss of rare navigation space by conservative approximations of collision shapes. The algorithm employs a low-dimensional sampling space to generate inherently drivable candidate motions. The sampling space is hereby designed to yield expressive candidate motions for the task of navigation on a road network. The planning and collision detection modules are both extensively tested on a self-driving vehicle prototype operating in mixed traffic on several indoor parking garages and outdoor parking lots. Hereby, the automated vehicle is solely equipped with low-cost sensors such as cameras and ultrasonic sensors. The motion planning approach above treats other agents in a non-cooperative fashion. First, predictions of the future motion of objects are obtained. Afterwards, motion planning is casted to the problem of finding a collision-free trajectory with respect to the fixed object predictions. It is shown from statistics of extensive real-world experiments, that---despite good performance in static environments and selected dynamic scenarios---the non-cooperative approach is especially vulnerable to noise in velocity estimates of other agents when navigating in close proximity. In the absence of rich semantic information to stabilize long-term object predictions, this thesis advocates the assumption of cooperative behavior as an autarkic mechanism to alleviate over-conservative robot behavior in close proximity to other agents. A cooperative probabilistic planning approach based on a maximum entropy model of joint navigation behavior is pursued, where typical per agent motion patterns, as well as joint navigation patterns are captured via feature functions. Parameters of the model are learned from demonstrated behavior in order to circumvent the cumbersome and difficult process of manual parameter tuning. Motion planning with the probabilistic model results in a joint trajectory optimization problem, in which the most likely collaborative solution is sought. As the model includes feature functions for static obstacle avoidance of all agents as well as features for agent interaction compliant with social norms, complex joint navigation scenarios can be forecasted. This thesis uniquely demonstrates real-world cooperative planning with the maximum entropy model in an uncontrolled public space. The results obtained support that cooperative planning is mandatory to address over-conservative mobile robot behavior in environments shared with other traffic participants Show more
External linksSearch via SFX
ContributorsSupervisor: Siegwart, Roland Y.
Supervisor: Krause, Andreas
SubjectMotion planning; Self-driving vehicles; Object prediction; Cooperative planning; Autonomous navigation
Organisational unit03737 - Siegwart, Roland Y.
269916 - 60c303d6f8bb1017783566f49bfe12e5 (EC)
610603 - 4db57f7a2c84e3784bf20a2131c60216 (EC)
MoreShow all metadata