UC Berkeley

Historically, robots have successfully performed various tasks in isolated areas by following preprogrammed commands. However, more and more potential robotic applications require robots to complete tasks alongside or in collaboration with other agents such as other robots and human workers. Such robots need algorithms that enable flexible behaviors in crowded and uncertain environments, e.g., to plan actions safely and efficiently in shared spaces where other agents are present. Consequently, researchers have proposed modularized robotic systems, which typically consist of perception, prediction, planning, and control modules. However, most of these works only emphasize the design of a single module and use off-the-shelf methods directly for other modules. Therefore, the improvements made by the re-designed module may have limited value to the overall robotic system; the off-the-shelf methods may not be able to provide adequate information or fully utilize the information provided by the re-designed module. Robot modules should be designed for better performance individually and collectively; thus, this dissertation aims to develop integrated designs and strategies for modularized robotic systems in uncertain environments. Furthermore, the computational efficiency of the robotic system is crucial to the robot's performance in uncertain environments; we address the improvement of the computational speed of the most time-consuming module, the planner.

This dissertation consists of two parts. Part I involves a fundamental exploration of robot motion planners and proposes hybrid motion planners that combine and utilize different planning methods for better computational speed and plan quality, such as travel distance. Chapter 2 combines a sampling-based algorithm, RRT*, and an optimization-based algorithm, the convex feasible set algorithm (CFS). Chapter 3 focuses on long-horizon planning problems; it combines RRT*, CFS, the Interior Point OPTimizer, and segmented trajectory optimization. Chapter 4 focuses on motion planning for articulated vehicles; it combines a search-based algorithm, improved A-search guided tree, and utilizes results from reinforcement learning to guide the search. Simulation results demonstrate the advantage of the proposed motion planners in terms of computational speed and plan quality in static and deterministic environments cluttered with obstacles. Chapter 5 studies motion planning in dynamic environments and presents a hierarchical receding horizon control (HRHC) framework. The HRHC coordinates a motion planner, such as the planners presented in Chapters 2-4, with a safety controller to achieve safe and efficient robot motion in uncertain environments in simulations and real-world experiments.

Part II presents application-oriented integrated robotic systems that coordinate the predictor and the planner to make an effective and safe plan. Chapter 6 discusses the close relationship between the prediction and planning modules and identifies several conditions for realizing safe model predictive control in dynamic and uncertain environments; we present a predictor designed for better closed-loop robot performance. Simulations and real-world experiments that involve a robot working alongside a human worker are conducted; the robot can navigate safely in the presence of unexpected human movements. In Chapter 7, simulations with a computer assembly setting that involve a robot collaborating with a human worker are conducted; the proposed robotic system coordinates the prediction and planning modules to utilize human motion prediction and uncertainty estimation for robust task planning. The robot can generate time-efficient task plans when the human worker performs inefficiently. In Chapter 8, simulations involving an autonomous vehicle navigating in a parking lot while avoiding collisions with static and moving obstacles are conducted; the proposed system includes a hybrid environment predictor that makes short-term and long-term predictions of the surroundings and a strategic motion planner that reacts to the environment according to the predictions. The robot demonstrates the effectiveness of the proposed method in terms of motion prediction, safe tracking, retreating in an emergency, and trajectory repairing.