Goal Babbling

Goal Babbling is a recently introduced method for direct learning of the inverse kinematics within few hundred movements even in high-dimensional sensorimotor spaces. The inverse model can be learned form scratch without any prior- or expert-knowledge, even for a soft elephant trunk robot with unknown kinematic function [1].
This algorithm is inspired by infants’ early goal-directed movements, for example their way of learning to reach, and thus "learn to reach by trying to reach".
Goal Babbling shows high scalability without substantial extra cost in high dimensions up to 50 DoF for planar arms [2] and 9 DoF for the upper body of our compliant humanoid robot (COmpliant huMANoid - COMAN) [3].


Direction Sampling:
Direction Sampling is an approach can bootstrap coordination skills using Goal Babbling without pre-specifying goals, instead, the goals will be generated along a randomly chosen direction, which allows to discover the reachable workspace and learn the inverse kinematic mapping simultaneously.
Direction Sampling was previously developed for a 2D workspace with a simple planar arm model [4], whereas we scale it to full 3D workspace and the complex 9-DOF upper body of our humanoid robot (COmpliant huMANoid - COMAN) integrating simplified walking behavior by means of a simulated robot-floating base [3].

Learning Inverse Statics Models and Symmetries:
Goal Babbling has been modified to learn inverse statics models online and from scratch without using any controller. It has been tested for a 2 DoF planar arm, 3 DoF simplified human arm model, and 4 DoF COMAN arm. The results are very promising. Moreover, the efficiency of exploration has been increased by using the inverse statics characteristics "Symmetries" [5].