Neural Networks: An Overview - Neural Networks: Methodology and Applications - page 75

1.4.12 An Application in Mobile Robotics

Process control is the science that determines the control actions to which
a process must be submitted in order to guarantee that it will operate in a
prescribed fashion, in spite of unmeasured and unpredictable disturbances.
As an example of the role that neural networks can play in mobile robotics,
we describe the automatic control of a 4WD Mercedes vehicle, which was
equipped by the French company SAGEM with sensors and actuators that are
necessary for making the vehicle autonomous. Controlling the vehicle consists
in sending the appropriate signals to the actuators of the steering wheel, the
throttle, and the brakes, in order to keep the vehicle on a prescribed trajectory
with a prescribed velocity profile, in spite of disturbances such as wind gusts,
sliding, slopes, etc.
Neural networks are good candidates as ingredients in nonlinear process
control systems. They can implement any (su ciently regular) nonlinear func-
tion. As a result, they can be useful in two different ways,

• as models of the process, since the design of a control system generally

requires the availability of a model; neural networks are particularly useful
in internal model control, as described in Chap. 5;

• as controllers, i.e., for computing the control signals (e.g., the angle by

which the driving wheel must turn, and the velocity at which it should
turn) from the setpoint (e.g., the desired heading of the vehicle).

The vehicle that was controlled was a 4 wheel-drive vehicle equipped with
actuators (electric motor for actuating the driving wheel, hydraulic actuator
for brakes, electric motor for actuating the throttle) and two categories of
sensors,

• sensors that measure the state of the vehicle (proprioceptive sensors):

odometers on the wheels, angular sensors for the driving wheel and for
the throttle, pressure sensor in the brake system;

• sensors that measure the position of the vehicle with respect to the universe

(exteroceptive sensors): an inertial platform.

The navigation and piloting system is made of the following elements:

• a planning module, which determines the desired trajectory of the vehicle,

and its velocity profile, given the start and arrival points and the existing
roads;

• a guiding module, which computes the heading and speed setpoint se-

quences;

• a piloting module, which computes the desired positions of the actuators;

• a control module of the actuators.

In that structure, neural networks are present at the level of the piloting
module, where they compute the desired position sequences of the actuators
as a function of the heading and speed setpoint sequences [Rivals 1994, 1995].