Adaptive learning of common motion patterns for human aware robot navigation
- Publication Type:
- Issue Date:
NO FULL TEXT AVAILABLE. Access is restricted indefinitely. ----- Human Robot Interaction (HRI) is a field of growing importance with an increasing number of researchers engaging in this exciting area of inquest. Aging populations and rising cost of labor are driving factors in the development of HRI as it is believed that robots can soon be co-workers and companions not only in elderly care but in many different applications. As a caretaker, a robot could support nurses or be a companion to support a person in need over long periods of time. A robot as a coworker could be doing some jobs in an office environment such as delivering internal mail or cleaning. In either application, it could be argued that interactions between a robot and people may be enhanced by adaptive human oriented robot behavior. This, however, requires a robot to have the necessary prior knowledge with regard to human social behavior. Furthermore, it needs to be adaptive to ever changing behavior patterns. This thesis deals with the learning of common human motion patterns in populated environments and the use of that knowledge for improved robot navigation in the context of HRI. Human motion in a given environment is a complex result of social, psychological and physiological requirements and constraints. From this, the existence of common motion patterns, i.e. commonly taken paths, arises. Therefore, these motion patterns contain some information about human social behavior with respect to sharing spaces. However, there are significant challenges which need careful attention. Human motion is inherently variable, where small changes in an environment can lead to great changes in motion patterns. Therefore, learning should be on-going and adaptive. Ideally, learning should happen on-line and the model should be generalized enough to incorporate changes. Furthermore, in mobile robotics, limited observability is a challenging problem as usually a robot can only observe a small part of its environment at any given time. Hence, partial observations at different times have to be combined in learning. The major contributions of this thesis are the formulation of a novel probabilistic learning method for the learning and the representation of human motion patterns. The method is formulated for on-line learning with a robot's on-board sensors. The use of Sampled Hidden Markov Models (SHMM) enables the incorporation of new information while making observations. Furthermore, applications of such a learned model are presented, especially in the areas of people tracking, localization and path planning. Particle filter based people tracking is probabilistically fused with learned motion patterns, which yields significant improvements in the tracking performance as well as faster convergence in model learning. Ambiguities in the robot localization problem due to symmetric environments are effectively solved by incorporating the knowledge of learned motion patterns. Path planning is given a different flavor of social awareness with learned human behaviors. Results are presented both in simulation and using real world experiments carried out in an office environment using a robot equipped with a laser range scanner. The results prove the viability and effectiveness of the proposed approach.
Please use this identifier to cite or link to this item: