A trivial method of learning control

摘要

The tracking error along a repeated trajectory can be minimized by storing the smoothed feedback torque in a table, and using it as feedfoward torque in the next run. It results a simple learning controller which requires almost no additional computation in comparison to a joint servo control. In this paper the convergence, stability and robustness of this scheme are tested experimentally, and it is found how to choose its parameters. The control performance after learning is similar or better than with a nonlinear controller using a dynamic equation for compensating the arm dynamics.