Retuning Control Systems


Within the domain of actively balanced legged locomotion, it is necessary to re-tune control systems to reflect changes in the simulation or the desired behavior. We have implemented a tuning algorithm which will re-tune an existing control system to take into account a change in model or desired behavior. Automatically tuning these control systems is challenging because for most sets of control parameters it is difficult for a computer to tell what changes will improve the behavior. We make the simplifying assumptions that there is an area where the control parameters result in good behavior and this area moves slowly as the simulation changes.

To go from one simulation to another, we go as far as we can and still get reasonable running, use a modified downhill simplex to re-tune to get better running, and repeat until we have the desired final simulation. The example here is one where the amount of weight on the quadruped changes from 5kg to 200kg. In each mpeg, the quadrupeds are using the same controller, tuned identically. As a result, in the early mpegs only the quadrupeds with very little weight run. However, as the tuning algorithm progresses more heavily laden quadrupeds begin running.

Video

Publications


Simulations HomeBack to the Leg Lab Home Page
© 1995 MIT Leg Laboratory. All Rights Reserved.
545 Technology Square, Rm. 006
Most recent update: 11/16/95