Human-like robot navigation in human-living environments

Abstract

With the advancement of robot navigation technology, robots are entering people's daily life from previously confined and enclosed workspaces such as laboratories and factories. However, most mobile robots can only navigate in limited, closed, and controllable scenarios. Navigating around the open human-living environments remains challenging. This thesis studies human-like navigation algorithms for small and low-speed mobile robots in open human-living environments. This thesis first presents a learning-based multi-agent collision avoidance method to navigate robots in dynamic scenarios. This method can be trained only in simulation and deployed directly to robots in real-world environments. Navigation in open environments needs to deal with a large number of unseen and diverse scenes, which poses a challenge to the adaptability of robots' navigation behaviors. Therefore, we propose an uncertainty-aware resilient navigation policy, which can adaptively adjust the navigation behaviors according to the uncertainty. To investigate a human-like tightly coupled perception and planning framework, we propose a planning-assisted perception framework designed to help robots recover localization when navigating dense crowds. Furthermore, to achieve the human-like capability of online understanding scenes, we first introduce the online dynamic object removal algorithm, which can remove the interference of dynamic objects and accurately model the spatial structure of the scene. Besides, for moving objects, the motion flow of dynamic objects was proposed to understand the traversable area and social preference for the scene. The above methods have been validated in simulated and real-world human-living environments.