Adaptive obstacle avoidance using residual HJB corrections

An algorithm for learning transition cost estimates for obstacle avoidance and path planning in two-dimensional analog-valued obstacle fields is presented. The approximate transition costs are modeled with CMAC (cerebellar model articulated controller) neural architectures. Training sets are generated via residual transition cost model corrections obtained from the principle of optimality equations of dynamic programming, a discrete version of the Hamilton-Jacobi-Bellman (HJB) equation of optimal control. Two obstacle field resolutions, one fine and one coarse, are used to derive the cost updates. The set of updates then provides the next generation of training data to the neural architectures