Forward-only Simulation for Agent Search

We present work on optimization search in a domain where random restarts are unavailable, a property of many real-world problems. In particular, we present a new technique for planning and control of agents that have high degree-of-freedom manipulators such as octopus tentacles. The specific type of manipulator that we investigate is known as a muscular hydrostat. A challenge for creating agents that use such manipulators is finding appropriate control parameters to move the end effector to a goal position in a cluttered environment. We physically model hydrostats as collections of masses, springs, and muscles, and perform goal-seeking through optimization over muscle parameters. We introduce a method of optimization called minima filling, a general approach that successfully finds a global minimum without the use of random restarts and through only incremental, stochastic movements. Using minima filling as a base, we construct probabilistic roadmaps (PRMs) and use them to navigate a simulated muscular hydrostat around obstacles in cluttered environments.