Autonomous Vision-based Docking of a Mobile Robot with Four Omnidirectional Wheels

Sammanfattning: Docking of mobile robots requires precise position measurements in relation to the docking platform to accomplish the task successfully. Besides, the pose estimation of the robot with the sensors in an indoor environment should be accurate enough for localization and navigation toward the docking platform. However, the sensor measurements are entitled to disturbances and measurement uncertainties. In this thesis, sensor integration is exploited to decrease the measurement errors and increase the accuracy of the docking. The first approach in the autonomous docking of a mobile robot considered in this thesis is to use the already built-in laser scanner sensor to examine the feasibility of the precise docking with several experiments employing different markers. The second approach is a vision-based control method with a marker mounted on the docking platform, identified as the target. The vision-feedback control system evaluates the current position of the robot relative to the target in real time to compute the velocity commands for the actuators to reach the docking platform. The versatility of the markers is investigated in this method. The final approach is a Reinforcement Learning (RL) framework to investigate and compare the optimality of the docking with the vision-based control method. In the RL approach training is handled in a simulation environment with the reward distribution. The obtained results of the approaches are evaluated in experiments based on the desired docking behavior with respect to the trajectory and time. The control design with a real-time vision system demonstrates the capabilities of this approach to conduct accurate docking of the mobile robot starting in different configurations.