Is Growing Good for Learning?

In both living organisms and human-made machines, we are interested in how morphology (or mechanical design) and neural control (or control) are combined to reach a desired behavior. Engineers tend to split the mechanical and controller design into two separate steps, since it is difficult to analyze the effect of a controller without already having model for the mechanical dynamics. In recent years, control engineers have tried more and more to exploit the natural dynamics: this means, the controller should not overpower what the passive mechanical system naturally tends to do, but rather simply push it in the right direction then let the natural dynamics take over. This concept was pioneered in the engineering field by Tad McGeer[5], and also shown to be prevalent in biology[3][7]. This concept highlights the importance of mechanical design on controller design[6], especially for controller-design algorithms such as reinforcement learning where stable natural dynamics can provide a smooth gentle slope to help the algorithm to converge[9].