Understanding Sensori-motor Coordination during a Humanoid Robot Dynamic Task

A coordination control structure is employed to manipulate the dynamics of a simulated 3D articulated mobile robot. Coordination emerges as the interaction between robot joints, its parameters are computed using a policy gradient reinforcement learning process, obtaining a machine extrapolation of human cognitive behavior: The interplay of brain, body and environment. A simulated Hoap-2 humanoid is trained to achieve an equilibrium task through experience. The system learns how to change the dynamics of the articulations to assess coordination, avoiding to fall down during the execution of the task. The robot is able to automatically compute their motions from high-level descriptions of tasks, without the use of pre-established models, but with ones acquired by sensing.