Sensory and Memory-Based Path-Planning in the Egocentric Reference frame of an Autonomous Mobile Robot

This paper describes a path planning system using combined information from memory and sensors in a egocentric map. This technique enables autonomous mobile robots to use optimal path planning techniques such as the resistive grid. Using an egocentric map solves the self-localisation problem for mobiles robots and facilitates the reading of potential gradients in the resistive grid because the central node corresponds always to the position of the robot. In the described system, obstacles found with distance sensors are placed on the portion of an egocentric sensory map corresponding to the "visible space", with angles and distances determined by the current orientation and position of the robot. This current sensory knowledge is used to retrieve the positions of other, currently invisible, obstacles from an invariant spatial memory. The knowledge from the memory is then added to the sensory knowledge on the sensory map which has short-term memory nodes connected one-to-one to those of the resistive grid. The activities of nodes in the sensory map constrain the activities of the nodes in the resistive grid corresponding to target and obstacles positions, thereby providing the necessary information for path planning. After each movement of the robot, the content of the sensory map is reconstructed. We have developed a form of invariant spatial memory based on the recognition and storage of views of the obstacles configuration observed from a set of focus points. These are objects chosen for their special appearance, for instance corners. Such views are stored in a polar representation. For their recognition by "view-neurons", a polar shifting map has been developed. The behaviour of the view-neurons connected to the shifting map is reminiscent of that of place cells in the hippocampus. This form of invariant spatial memory allows to learn a number of different new environments of any size and to retrieve relevant information when required. The proposed system is demonstrated by simulations. Possible and necessary improvements are discussed. The system can be extended to path planning in states spaces of any dimensionality and offers the potential for mental exploration based only on information from memory placed in the sensory map.