Probabilistic Motion Planning in Uncertain and Dynamic Environments

Probabilistic Motion Planning in Uncertain and Dynamic Environments

Zhou, Bingyu (TU Delft Mechanical, Maritime and Materials Engineering; TU Delft Delft Center for Systems and Control)

Alonso Mora, J. (mentor)

Delft University of Technology

Mechanical Engineering | Systems and Control

2017-09-19

Autonomous driving is one of the popular and advanced research fields aiming to reduce the mortality rate and improve the welfare and efficiency of commuters' lives. As one of the important research branches of self-driving cars, this thesis focuses on the motion planning problems in dynamic and uncertain environments. The challenge of motion planning in uncertain and dynamic environments is that the ego-vehicle or the robot needs to intelligently reason about the interactions between itself and the other traffic participants. In order to successfully make decisions to guarantee the safety and users' comfort, it requires the motion planner to account for the future motions and the uncertainty of motion intentions associated with the obstacles to avoid the ego-vehicle ending up with inevitable collision states. This thesis presents three approaches to solve the challenge from manifold perspectives. The first approach, multipolicy MPC, introduces the uncertainty of motion intentions into the traditional optimization framework by modeling the multiple hypotheses of future trajectories as mixture Gaussian distributions. The second approach, centralized MPC, computes the motion plans for all vehicles including the obstacle vehicles by assuming the obstacles are controllable and broadcast their motion intentions. The last approach called joint behavior estimation and planning is a novel algorithm, which takes the interactions into account. It leverages the strengths of online POMDP to model the interactions through anticipating the future trajectories of obstacles under different motion intentions. The predicted trajectories are then utilized in the multipolicy MPC motion planner to compute the optimal actions for the ego-vehicle. A parallelized structure of joint behavior estimation and planning is also presented to scale up in cluttered environments. Simulation results demonstrate the benefits of the proposed approaches, particularly the joint behavior estimation and planning, in uncertain and interactive environments.

Motion planning
Trajectory generation
Interaction
MPC
POMDP

http://resolver.tudelft.nl/uuid:f491f7d8-a2f5-4f89-b4b7-86ac6b64546b

2018-09-19

Student theses

master thesis

© 2017 Bingyu Zhou