Link to refer to this document:
Abstract
Keyframed motion planning is a technique that specifies a robot motion by its joint variable samples in discrete time-steps. In this paper, we aim to provide an off-line (i.e. non real-time) dynamic motion optimizing method for keyframed humanoids. Let´s assume that a desired reference movement has been designed, it can be simulated using a real-time kinematics model. Due to dynamic effects the robot segments will not exactly follow the reference trajectories. Assuming a detailed, sophisticated dynamics model (running offline) we can formulate a norm that expresses the difference of dynamic and kinematic simulations. In this article we present our idea, how the motion could be automatically tailored by lowering this norm using numerical methods in a way, that the output of the dynamic model better approximates the reference motion. Finally, we show our experimental results within a modern simulation environment as well as on our test humanoid platform.
