Dynamic Motion Learning for a Flexible-Joint Robot using Active-Passive Motor Babbling

Dynamic motion taking advantage of inertia with a flexible-joint robot is useful for energy efficiency and rapid motions. However, it is difficult to control flexible joints considering the complexity of their dynamics. To overcome the problem in past studies, oscillator [1], attractor [2], and search tree [3] methods have been explored. Attractors and oscillators force trajectories to return to the pre-designed trajectories when any deviation occurs, while the search tree finds suboptimal motion. These methods have three limitations: (1) Generate only designed motion; (2) require all robot model information; and (3) take time to search. To overcome these problems, we use recurrent neural network (RNN) and motor babbling to acquire body dynamics, forward inverse model [4]. RNNs can learn various motions and therefore, robots will be able to acquire body dynamics with motor babbling. Motor babbling is defined as the movement that infants use to acquire their own body model [5]. RNNs can also generate associative motions, which do not take time to search. In our previous research [4], a robot performed simple motions using motor babbling to acquire body dynamics with RNN. However, to generate more complex motions, it is necessary to learn specific movements. It would take a long time to learn such specific tasks with neural networks. Therefore, we propose a method where the robot learns specific tasks after the robot performs pre-training with motor babbling to acquire body dynamics. To acquire body dynamics (natural dynamics) with a flexiblejoint robot, we also target the type of motions used in motor babbling.