AUTONOMOUS MOBILE ROBOT NAVIGATION IN DYNAMIC ENVIRONMENTS

Abstract

In autonomous navigation for mobile robot in dynamic environments, there are multiple challenges. For instance, localisation is prone to error while obstacle avoidance has to accommodate the trajectories of dynamic objects. This thesis proposes 2 different approaches to independently solve localisation and obstacle avoidance problem. To improve localisation accuracy, ALOT (Augmented Localization with Obstacle Tracking) is proposed to utilise dynamic object tracking with particle filter in state estimation. The object tracker provides the robot’s ego-pose proposal in particle weighing schemes. It has been shown that ALOT achieves lower erros than the popular AMCL in localisation in highly dynamic environments. Additionally, the Hybrid Dynamic Window Approach (HDWA) with ExpMap is developed to improve obstacle avoidance in dynamic environments by leveraging data-driven approach. ExpMap utilises the deep neural network to encode the robot’s past control experiences and occupancy prediction from RGB image and state’s input. These experiences are then combined with the physics-based approach of the popular Dynamic Window Approach controller by incorporating 2 additional cost influenced by the ExpMap predictions. With this, HDWA was able to leverage its past experiences while ensuring safety. It has been demonstrated that HDWA was capable of driving the robot slowly near people and small gaps which interestingly include glass doors.