Erosion of basins of attraction: performance losses in sensorimotor learning of a robot manipulator

The sensorimotor learning phase is considered in the context of a collision-free dynamic path planning architecture for a robot manipulator in an uncertain task environment. The robot dynamics are effected by a vector field generated by partially known attractors and repellers, with uncertainty represented by entropy measures. The joint space is transformed into a cell space, and sensor uncertainty is taken proportional to cell size. The authors use a cell-to-cell mapping concept to develop a simple cell mapping algorithm to find the basin of attraction of each attractor cell. These basins constitute the valleys in cell space in which the training vectors should lie. Learning predictability is considered to reside in the nonfractal properties and width of basins of attraction, and predictability loss measures are developed from basin erosion under parameter changes