A connectionist approach to autonomous robotic navigation
Weiland, Peter Lawrence
Cheatham, John B., Jr.
Doctor of Philosophy
Robotic navigation has been an area of intense research since the onset of mobile robot development. The usefulness of mobile robots ultimately reside in their ability to move and interact with the environment. Current approaches to robotic navigation are primarily based on simulating intelligent, human-like behavior through the intelligent system model processing cycle; sense, perceive, reason, act. Unlike these methods, this thesis presents a navigation system based on biological and behavioral principles which functions in a stimulus-response manner. Using connectionist architectures, a relationship between stimulus and response is required through the learning of conceptual information pertaining to navigation. In this research, the mammalian visual system provides a guide for the processing of environmental stimulus. Simulated laser range data are processed in retinal patch size elements by a cellular neural network. This network is designed to detect obstacle existence for each patch segment based on an invariant feature of range discontinuity. Obstacle information is then mapped in binary format, indicating the traversable state of the patch, to the system's visual cortex. Response to this mapping is derived from a hierarchical structure of back error propagation neural networks in which each network has learned a particular navigational behavior; obstacle avoidance, wander, and goal seeking. Output from these networks indicate an appropriate motor response for the environmental stimulus. A simulation was developed to evaluate the performance of this system by having a robot traverse an environment. The connectionist approach was verified through system display of human-like navigational behavior for the simulation's environment. Advantages of the neural network approach were also demonstrated by its processing speed and adaptability. Procedures are discussed for actual system implementation in which cycle times of under one second are completely feasible. Proposals for unsupervised learning of navigational responses for environmental stimulus are also made. From the research presented in this thesis, a foundation is established for continuing the study of the connectionist approach to the problem of autonomous navigation.
Mechanical engineering; Computer science; Artificial intelligence