Re: Learning control of robots

Mike Fouche (
Fri, 15 Mar 1996 19:18:30 +0100


If I understand you correctly, your saying that the neural network is
trained on-line using the error as the difference between output of
the conventional controller and the neural controller itself.

Not to dampen your enthusiasm but this is a common technique which
people have mentioned to me before and I have seen in the literature.
HOWEVER .. that doesn't make your work invalid or less important! If
you can gain experience doing this in hardware there are a lot of
applications that can benefit from this. Remember that "everybody and
their grandmother" has new ideas but what really is important is
getting the stuff implemented into hardware and fielded. That's where
MANY lessons are learned. This is what separates the men from the

We are also doing nonlinear control with neural networks for several
applications. In terms of robotics, we're only at the simulation
stage. However we have a technique in which the settling time and
required force/torque can be attenuated by the neural controller. The
technique proved invaluable on a cart/pendulum hardware system in
which the physical properties of the system had changed over time.
The classical controller generated high frequency behavior that was
nearly unstable. The neural controller, which was trained from the
classical, was able to "phase-out" the high frequency behavior and
large amplitudes.

Anyway - my unsolicited advice to you is ... don't worry too much
about what others are doing. Just become good at the hardware
implementation of neural control. The market is getting bigger - the
hardware experts will be in demand.

Mike Fouche
Boeing Missile & Space Division

*****   Crossposted with comp.robotics.research (moderated)  *****
  Summary: Academic, government & industry research in robotics.