Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)

Graduate Group

Electrical & Systems Engineering

First Advisor

Vijay Kumar


Stochastic motion planning is of crucial importance in robotic applications not only because of the imperfect models for robot dynamics and sensing but also the potentially unknown environment. Due to efficiency considerations, practical methods often introduce additional assumptions or heuristics, like the use of separation theorem, into the solution. However, there are intrinsic limitations of practical frameworks that prevent further improving reliability and robustness of the system, which cannot be addressed with minor tweaks. Therefore, it is necessary to develop theoretically justified solutions to stochastic motion planning problems. Despite the challenges in developing such solutions, the reward is unparalleled due to their wide impact on a majority of, if not all, robotic applications. The overall goal of this dissertation is to develop solutions for stochastic motion planning problems with theoretical justifications and demonstrate their superior performance in real world applications.

In the first part of this dissertation, we model the stochastic motion planning problem as Partially Observable Markov Decision Processes (POMDP) and propose two solutions featuring different optimization regimes trading off model generality and efficiency. The first is a gradient-based solution based on iterative Linear Quadratic Gaussian (iLQG) assuming explicit model formulations and Gaussian noises. The special structure of the problem allows a time-varying affine policy to be solved offline and leads to efficient online usage. The proposed algorithm addresses limitations of previous works on iLQG in working with nondifferentiable system models and sparse informative measurements. The second solution is a sampled-based general POMDP solver assuming mild conditions on the control space and measurement models. The generality of the problem formulation promises wide applications of the algorithm. The proposed solution addresses the degeneracy issue of Monte Carlo tree search when applied to continuous POMDPs, especially for systems with continuous measurement space. Through theoretical analysis, we show that the proposed algorithm is a valid Monte Carlo control algorithm alternating unbiased policy evaluation and policy improvement.

In the second part of this dissertation, we apply the proposed solutions to different robotic applications where the dominant uncertainty either comes from the robot itself or external environment. We first consider the the application of mobile robot navigation in known environment where the major sources of uncertainties are the robot dynamical and sensing noises. Although the problem is widely studied, few work has applied POMDP solutions to the application. By demonstrating the superior performance of proposed solutions on such a familiar application, the importance of stochastic motion planning may be better appreciated by the robotics community. We also apply the proposed solutions to autonomous driving where the dominant uncertainty comes from the external environment, i.e. the unknown behavior of human drivers.In this work, we propose a data-driven model for the stochastic traffic dynamics where we explicitly model the intention of human drivers. To our best knowledge, this is the first work that applies POMDP solutions to data-driven traffic models. Through simulations, we show the proposed solutions are able to develop high-level intelligent behaviors and outperform other similar methods that also consider uncertainties in the autonomous driving application.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Additional Files

failure_case_tree_tangle.mp4 (18438 kB)
failure_case_warehouse_collision.mp4 (11955 kB)
pomcp++_single_mode_initial_belief.mp4 (4649 kB)
pomcp++_two_mode_initial_belief.mp4 (4795 kB)
RTGNN.mp4 (27740 kB)