Reaching through learned forward model

This paper presents a learning approach for a humanoid to reach objects in its environment. Instead of assuming that the exact forward kinematics of the arm is given, we address the reaching problem by first learning forward kinematics with a RBFN through autonomously gathered training samples. The learnt forward model is subsequently used to construct Jacobian matrices to incrementally generate straight reaching trajectory exhibited by humans. We show that if the learning parameters are set appropriately, a RBFN trained on a small number of samples corrupted by perception noise can still lead to high reaching accuracy. The size of the training set can be further reduced without severe performance degradation if limited visual feedback is used to aid reaching after the end effector has been moved into the neighborhood of the desired object.