A Monte-Carlo Algorithm for Path Planning With Many Degrees

We describe a stochastic techuique for p l d n g colllion-frec path. of robots with many deof freedom (DOF's). The algorithm incrementally builds a graph connecting the local "a of a potential function defined in the robot's configuration space and concurrently searches this graph until a goal configurationis attained. A local minimum is connected to another one by executing a random motion that hopefully escapes the well of the first minimum, succeeded by a gradient motion that follows the negated gradient of the potential function. All the motions are executed in a grid thrown through the robotC configuration space. The random motions an implemented M random walks which an known to converge toward Brownian motions when the steps of the walks tend toward zero. The 1 0 4 minima graph is searched using a depth-first strategy with random backtracking. This technique avoids the planner to explicitly represent the local-minima graph. The path planning algorithm has been fully implemented and run successfully on a variety of problems involving robots with many deof freedom. T h e problems are far beyond the practical capabilities of previously existing plannera. In particular, our planner w~ able to generate path. for a 31-DOF robot arm in a 3D Acknowledgements: This r e d WM funded by DARPA contract DAAA21-89C0002 (Army), DARPA contract N00014-SK0620 (Office of Naval Fiewar&), CIFE (Center for Integrated Fscility Engineering), CIS (Center of Integrated Systems), and DEC (Digital Equipment Corporation). The authom thank Ilan Caron for hie work on the path smoother. workBpace.