Neural Architecture for Robot Planning

Abstract

Human motor skills depend, to a large extent, on two factors: (i) the intrinsic compliance of the musculo-skeletal system and (ii) its kinematic redundancy. The former factor allows a smooth modulation of contact forces during manipulation as well as a partial compensation of disturbances during trajectory formation. The latter is necessary for incorporating different task constraints in the same motor plan. In particular, experimental investigations of human arm movements [3] showed that the arm (even of a deafferented monkey, deprived of any kinesthetic/somesthetic feedback) returned toward an intermediate position (between the initial and final one) when the arm was temporarily displaced before the onset of the target. This suggests the hypothesis that the central nervous system plans a movement in terms of a sequence of equilibrium points, or virtual trajectory [5] but does not tell us anything about the central process which is responsible for producing it, particularly in the case of a redundant system. The main idea of this paper is that in complex robotic manipulators which, similarly to the human arms, are characterized by mechanical compliance and kinematic redundancy, it is quite useful for the planner to establish an analogy beween the real-time gradient-descent process determined by the mechanical potential field and a similar process associated with the dynamics of a neural computational engine which produces the virtual trajectories. In fact, the reaction of the arm to a contact/disturbance force is a gradient-descent in the elastic potential field of the musculo-skeletal system and this does not imply just a single motion pattern but a family of responses, indexed by the stiffness level of the different muscle groups, thus allowing a run-time adaptation of the reactive-part of the plan. Analogously, the run-time adaptation of the trajectory-formation part of a plan can be achieved by a gradient-descent process in a computational potential field modulated according to different task constraints. In this paper, we show how artificial potential fields can be expressed by means of a self-organized map [8], which also represents a forward model of the manipulator, and how gradient-descent can be performed in real-time on this map.