Continuous task-priority rearrangement during motion execution with a mixture of torque controllers

When controlling highly redundant humanoid robots multiple motion tasks can be accomplished simultaneously by employing several reactive torque-control modules. Flexible task-prioritization for these modules during motion generation is an open research problem. Null-space based approaches for redundant robots implement a strict, hierarchical prioritization and are not suited for changing the priority-order of the underlying control modules. In contrast, the mixture of controller approach superposes torques from several control modules and thereby enables flexible and continuous task-priority rearrangement. Here we show how to react to environmental changes by blending the corresponding priorities during motion execution. Additionally, we investigate the advantage of regularization when learning new mixture coefficients. The proposed framework is demonstrated for the full 11 DOF upper body of the compliant humanoid robot COMAN in simulation.