Humanoid robots have been promoted for being able to work in real environments such as extremely hazardous situations instead of a human. However, their agility is not comparable to that of human beings. While robotic capabilities are rapidly advancing, the robot operation is dozens of times slower than a human demonstration.
The low speed of task execution can be partially attributed to two factors: its control strategy with multiple high-level controllers and the lack of a strong actuator. We, therefore, aim to develop an ideal framework, where a unified whole-body controller of Model Predictive Control (MPC) plans motion trajectories under full-body dynamics of a humanoid robot and then the desired whole-body motions are executed with a strong actuation system: a hybrid actuator system.
In this thesis, heading towards the development of the framework, we address two problems: computational burden of MPC and torque distribution problem of the hybrid actuator system. To deal with them, we propose a hierarchical MPC method and a two-stage control scheme based on a singularly perturbed system of a humanoid robot and a hybrid actuator. In addition, we develop an estimation framework of task goals from human demonstrations because designing appropriate control objectives is crucial but labor-intensive issue to generate the robot motions with MPC.
We show that the proposed hierarchical MPC successfully reduces the computation time of MPC without significantly degrading the control performance. Furthermore, we demonstrate that our two-stage control scheme successfully finds the torque distribution strategy in real-time for a hybrid actuator system. Finally, it is revealed that control objectives of a humanoid robot can be learned from captured human movements with our estimation framework.