Techniques for Head Control in Robot Biped Walking

Analyzing the available literature of motion-captured head motion during human walking, a simplified reference model of human-like head movement for humanoid robots has been developed. Torso motion during the WABIAN-2R humanoid robot walk has been recorded using Inertial Measurement Units. Several controllers have been developed to compensate for torso motion and to follow arbitrary references: a linear controller and two non-linear ones, based on inverse differential kinematics and on closed-form inverse kinematics formulae. The controller designs have been tested with an approximate dynamic simulation of the robot head. The bio-inspired Recurrent Cerebellar Architecture neural network model has been applied to the problem of learning the head rotational inverse kinematics. Such a neural network, even with few neurons compared to the size of the input space, is able to learn to invert the kinematics to an acceptable degree of precision