Motion planning for non-holonomic mobile robots in obstacle-rich environments

Autonomous ground vehicles (AGVs) are playing an ever-increasing role in civilian, industrial and military fields. For accomplishing such missions in complex environments, the capability of fully autonomous navigation while avoiding unexpected collisions is a fundamental and crucial requirement. In this thesis, we concentrate on the non-holonomic motion planning problem, and present methodologies and system designs that enable mobile robots to efficiently navigate in obstacle-rich environments. As the first expedition in this direction, a hierarchical control scheme is proposed to solve the motion planning and control problem of AGVs using only onboard resources. A model predictive control (MPC)-based trajectory planning and tracking control is developed at the high level, which enables better integration between planning and control and therefore enhances the kinematic-level tracking performance. A reduced-order extended state observer (RESO)-based dynamic control is developed at the low level. The uncertainties in the system dynamics are estimated via a RESO and then compensated in the control in real-time. Simulations and experiments with different payloads demonstrate the capability of the AGV to navigate safely and optimally in complex environments with large dynamic uncertainty. Based on the above research, we further investigate the multi-robot motion coordination problem in obstacle-rich environments. A novel centralized multi-robot trajectory planning method is developed, which consists of front-end path searching and back-end nonlinear trajectory optimization. We adopt a multi-agent path searching (MAPF) algorithm to find collision-free time-optimal initial paths, which are then refined into smooth and dynamically feasible trajectories. Besides, a prioritized trajectory optimization method is proposed to improve the scalability of the back-end algorithm. Extensive simulations and experiments verify the effectiveness and superiority of the proposed method. While the centralized multi-robot motion coordination method performs well in static environments, it cannot handle dynamic obstacles in the environment in real-time. Thus, we propose a novel distributed multi-robot motion coordination method, which is based on the optimal path following control and online conflict resolution. Each robot finds its path to the goal and follows it optimally using model predictive contouring control (MPCC). Meanwhile, possible conflicts between robots are detected based on the communication network. A prioritized path-following method and an event-triggered path re-planning mechanism are proposed to resolve all detected conflicts online. We verify the effectiveness of the proposed method via real-time simulations in obstacle-rich and dynamic environments. The aforementioned methods all rely on a pre-built or online constructed map, which is not always available in practice. To solve the map-less navigation problem, we develop an end-to-end mobile robot navigation method that directly maps raw sensor data and goal information to control commands using deep reinforcement learning (DRL). The main challenge of the application of DRL-based methods is the difficulty of training and the convergence speed. Therefore, we decompose the tasks of goal reaching and collision avoidance and propose a behavior-based navigation framework. The two basic behaviors are obtained separately and fused based on an estimated risk level, which enables a fast learning process. Simulations and real-world experiments demonstrate that the proposed method enables collision-free autonomous navigation of multiple mobile robots in unknown environments.Doctor of Philosoph