Cognitive maps for mobile robots: A representation for mapping and navigation.

摘要

Autonomous mobile robots need to explore, map and navigate the environment in which they find themselves. Their task is made even harder by inaccurate sensors and motors, and the dynamics and complexity of the environment. People face many of the same obstacles when exploring: mapping and navigating their environments, yet are undoubtedly successful at it. Therefore, this thesis identifies those properties and functions that underly a person's internal representation of his or her spatial environment (called a cognitive map), and applies those properties and functions to a mobile robot mapping system. The resulting system is compared with existing mobile robot systems across several dimensions, including performance, safety and robustness. The starting point for this research is a cognitive model of human cognitive mapping called PLAN, which is adapted for implementation on a mobile robot; the resulting system (RPLAN) is tested in a natural, indoor environment. RPLAN consists of three components. The first systematically integrates sonar and vision sensing using a Bayesian network to perform place recognition. Experiments show how sonar and vision can work together to produce a system that achieves better results than if either are used separately. The second component of RPLAN selects a route to a goal and directions to the next place along that route. To do this RPLAN uses a novel spreading activation network that encodes direction and familiarity. Finally, RPLAN creates a global overview of its environment from its local representations. The overview is image-like and can help solve simple spatial problems such as finding shortcuts and detours. The overview can also be used to improve route selection. RPLAN is the only computational theory of cognitive mapping implemented, at least in part, on an actual mobile robot and introduces many new ideas to mobile robots from the cognitive mapping field, including image-like spatial representations, biased route networks and the integration of sonar and vision for place recognition. In the future, mobile robotics will benefit from closer ties with researchers in cognitive mapping and vice versa.