Learning Motor Control for Simulated Robot Arms

Controlling a high degree of freedom humanoid robot arm to be dextrous and compliant in its movements is a critical task in robot control. The dynamics of such flexible and light manipulators have a highly non-linear nature, making analytical closed form solutions using rigid body assumptions inappropriate. In this thesis, we use Locally Weighted Projection Regression to learn online the inverse dynamics of a high degree of freedom dextrous robot arm in a physically simulated environment. The learned control task is based on a visual servoing scenario, which incorporates trajectory planning, inverse kinematics and motor control. We build a powerful simulation framework for robot arms using the Open Dynamics Engine. The developed software places emphasis on flexible creation of any arbitrary robot arm, by incorporating polyhedral object geometry for optimised 3D rendering and physical parameter calculation. The extensible software architecture allows for easy incorporation of other learning algorithms and control paradigms. The goodness of the learned inverse dynamics and the simulation model is verified using several experiments. Finally we learn the inverse dynamics model of a robot arm for a defined operating region, and show the goodness of the learned model by evaluating it against recorded human motion. We aim in the long run to transfer the gained insights and results of the reference arm to a real hardware implementation. ii Acknowledgements