Walking despite the passive compliance: Techniques for using conventional pattern generators to control instrinsically compliant humanoid robots

There are several problems which arise when using a standard ZMP-based pattern generator to control an intrinsically compliant humanoid robot like COMAN. We present two techniques: pelvis forward trajectory smoothing and polynomial admittance gain modulation, which make it possible to use the conventional pattern generator to control such passively compliant robots. The former method modifies the reference of the pelvis trajectory to counteract its overshooting caused by the compliance of the legs. The latter method is meant to decrease the impact during initial contact and decrease the error between the foot position and original reference during the mid-stance caused by the use of admittance control. We explain details of both of the methods and show results from walking experiments with and without the controls, proving their effectiveness.