Neural networks for robot arm control
Static robot arm positioning
Patrick van der Smagt, Ben Kröse, Frans Groen
Over the last few years, a neural network was
designed and implemented to control a six DOF robot arm, equipped with
a camera in the end-effector. The task of the arm is to position the
end-effector at a small distance above the target. The controller is
able to learn the mapping between the camera image domain and the joint
domain of the robot, and performs the operation in 3D. Within this project
fast learning algorithms were developed and implemented,
so that the controller can, in real time, adapt to changes in the system.
Time-to-contact control of a monocular robot arm
Patrick van der Smagt, Ben Kröse, Frans Groen
Instead of using static visual info, it is also possible to use
time derivatives of the visual data. This is also used in biological
systems, such as the gannet.
When hunting for fish, this bird drops itself from the sky from a great
height. Since the fish is moving, the bird needs its wings to correct its
path while falling down; however, at the moment of contact with the
seawater its wings must be folded to prevent them from breaking. It has been
shown that the time remaining between the moment that the bird folds its wings,
and that it hits the water, is always the same for one particular
bird: the time-to-contact. We have shown that by extending
the above example to higher-order time derivatives
of the visual data, criteria can be developed which specify a
trajectory which ends in a rest state (i.e., zero
velocity and higher derivatives) at the end point.
These criteria will be the visual setpoints along
the followed trajectory. Thus it is possible that
the eye-in-hand robot arm exactly stops on an observed
object by use of optic flow.
Neural control of the dynamics of a rubbertuator arm
Patrick van der Smagt and Klaus Schulten
We have studied the trajectory control of a
pneumatically driven robot arm
resembing a skeletal muscle system. The arm dynamics have been
shown to be hysteretic and significantly changing in time due to
external influences thus requiring an adaptive controller.
Neural networks for target tracking
M.G.P. Bartholomeus, B.J.A. Kröse, A.J. Noest, G. Schram
Robot systems which use vision to track a moving target have
to 1) find a moving target with a moving camera, 2) determine the
correct mapping from the visual domain to the joint domain
We developed a vision system which is able to detect a moving target
in a moving background. A real-time processing is achieved by
implementation on a Datacube Maxvideo system.
For accurate control we developed a neural "predictive" controller,
using the observed velocity
of the target and the joint velocities as extra inputs in the network.
A tracking performance of about two centimeters for a moving target is
achieved.
