Learning Constrained Distributions of Robot Configurations With Generative Adversarial Network

In high dimensional robotic system, the manifold of valid configuration space often has a complex shape, especially under constraints such as end-effector orientation or static stability. We propose generative adversarial network approach to learn distribution robot configurations constraints. It can generate that are close constraint manifold. present two applications this method. First, by learning conditional with respect desired position, we do fast inverse kinematics even for very degrees freedom (DoF) systems. Then, use it samples in sampling-based constrained motion planning algorithms reduce necessary projection steps, speeding up computation. validate simulation using 7-DoF Panda manipulator and 28-DoF humanoid Talos.