Applying intrinsic motivation for visuomotor learning of robot arm motion

In this paper, we present a method to apply intrinsic motivation for improving visuomotor learning of robot´s arm with external object in view. Multiple Timescales Recurrent Neural Network (MTRNN) is utilized for learning the robot arm/external object dynamics. Training of MTRNN is done using the Back Propagation Through Time (BPTT) algorithm. BPTT algorithm is modified as follows. 1. Evaluate predictability of robot arm/objects using training error of MTRNN. 2. Assign a preference ratio to each object based on predictability. The preference ratio represents the weight of each object to training. Experiments were conducted using an actual robot moving the arm while a human moves his arm in the robot´s camera view. The result of the experiment showed that the proposed method presents better training result of robot arm visuomotor dynamics compared to general training with BPTT.