home        research        publications        people        open positions        contact
	Incremental Learning for Control Model-based control offers numerous advantages but requires accurate representation of the system dynamics. It is difficult to formulate an analytical model which fully captures all the complex nonlinear dynamics such as friction and backlash. Non-parametric regression techniques can be applied to approximate the inverse dynamics by observing the torques applied at the joints and the resulting motion of the system. Furthermore, time-varying parameters which may arise from joint wear and tear, can be accounted for by performing online and incremental updates to the data-based model. Various supervised learning algorithms such as Locally Weighted Projection Regression, Gaussian Process Regression, and Support Vector Regression are being investigated and tailored for use with robot manipulators.
			Here, the inverse dynamic model of a 6 degree-of-freedom robot is being learned while tracking a ‘Figure 8’ trajectory with a simple PD controller. The resulting model is subsequently used in an inverse dynamics controller to achieve high performance tracking. - video showing CRS A465 robot controlled with PD and LWPR control
	Researchers: Joseph Sun de la Cruz, Sheran Wiratunga, Rajan Gill, Bill Owen, Dana Kulić Key Publications: J. Sun de la Cruz, W. Owen and D. Kulić, Online Learning of Inverse Dynamics via Gaussian Process Regression, IEEE International Conference on Intelligent Robots and Systems, pp. 3583 - 3590, 2012. pdf J. Sun de la Cruz, E. Calisgan, D. Kulić, W. Owen and E. Croft, On-line Dynamic Model Learning for Manipulator Control, IFAC Symposium on Robot Control, pp. 794 - 799, 2012. pdf J. Sun de la Cruz, D. Kulić and W. Owen, A Comparison of Classical and Learning Controllers, World Congress of the International Federation of Automatic Control, pp. 1102-1107, 2011. pdf