Abstract:

From cold working machines to lovely electric pets, the robots are likely to continue to impact various aspects of our lives. In the near future, robots will consistently appear in human communities in schools, hospitals, offices, museums, and households etc. For robots to be socially accepted by humans, robots must have the ability to understand human behaviors and spatial relationships within environments. This dissertation attempts to develop the learning methods for spatial understanding of human society and the robust motion planning algorithms of mobile robots. New frameworks in different spatial representations are established. On the topological level, SLAMMOT-SP is introduced for simultaneous SLAMMOT and scene prediction. On the cognitive level, SBCM, PEG, and the concept of motion primitive learning are proposed to model generalized pedestrian behaviors. The probability models for spatial reasoning and behavior prediction are also derived. Moreover, several planning algorithms, DAO*, DDAO*, and predictive anytime A*, are presented to satisfy the requirements of anytime, fast replanning, and uncertainty concerns. Finally, we demonstrate that the robot is capable of predicting the intentions and long-term trajectories of pedestrians, and further behaving socially acceptable motions by combining planning algorithms with spatial reasoning.