Transcript grzeszczuk
Automated Learning of MuscleActuated Locomotion Through Control Abstraction Radek Grzeszczuk and Demetri Terzopoulos Presented by Johann Hukari • Animating animals is hard • Muscle interactions are complex, even in “lower” life forms • People notice when animals walk like they have a stick up their ass • Is it possible for a physics based simulation to control itself? • Real animals are efficient locomotors • Locomotion is a learned skill • Behaviors are constructed from lower level behaviors and combined into more complex actions • This paper explores extremely flexible bodies: Things that slither or swim • Bodies constructed from spring-mass systems, many DOFs • Providing animators with controls to every muscle (even simplified versions) is needlessly complex. • Instead of direct control, let the animal control itself • Generate and test. Did this change result in better motion? • Low level motions combined to generate more complex motion. • Learning low-level control can be lengthy, but is fairly simple • “Muscles” are control according to a scalar activation function • “Animals” are allowed time to learn simple locomotion and turns of various radii • These actions are combined together with approx 5% which is blended to disguise discontinuities • When given goals “animals” choose the locally optimal solution: “greedy” • Stupid pet tricks • Dolphin can learn “Sea World” style tricks by combining swimming and turning • Dolphin tail modeled by turning previously generated shark’s tail sideways. • Simply sequencing controllers greedily is simplistic • Other method: Learn complex behaviors by combining (rather than sequencing) • Doesn’t use intermediate solutions. • Complex combinatorial problem