The term ``action'' has been used widely in the ethology, psychology
and robotics literature, sometimes with quite different meanings. A
popular definition of action, in the context of action selection in
animals, is given by Tyrrell Tyrrell92a: an action refers to
one of the mutually exclusive entities at the level of the
behavioral final common path [McFarland and Sibly1975]. The level of the
behavioral final common path is the lowest level of control in an
animal or agent's control system, whereby all behaviors are
expressed. However meaningful and accurate from an ethologist's point
of view, this definition leads to a certain degree of confusion from a
designer's point of view. More precisely, it gives the impression that
an action is ``a movement'', hence the action selection mechanism in
an anthropomorphic robot, for example, is responsible for determining
the detailed movements of every muscle. As a result, one cannot help
wondering how the action selection mechanism is any different from the
whole control system of an agent. An alternative is therefore to
define an action as a motor skill, such as ``move forward'' or
``turn to the right''.
Depending on the mechanical model of an agent, each of its
actions may require a simple single movement of a single actuator or
muscle, or a complex series of coordinated movements of a number of
actuators.
Defining an action as a motor skill allows us to deal with the problem
of action selection at a higher level by differentiating mechanisms
for motor control from those for action selection. More
specifically, given this definition, motor control mechanisms are
responsible for controlling and coordinating actuators in an agent so
as to form useful motor skills, i.e. actions, while action selection
mechanisms are only responsible for choosing an action without knowing
how it is implemented. This concept underlies the design of the
artificial fish. We first build a motor control system to implement a
set of basic motor skills, including swim forward and backward, turn
right and left, glide, yaw, pitch, roll, and brake. The behavior
control system is then built to control the selection of these motor
skills in order to produce realistic behaviors.
Fig. 
illustrates how action selection differs from
motor control in a general design scheme. Note that exclusive
actions, i.e., actions that use the same actuators/muscles, cannot be
selected simultaneously. For example, one can not walk while sitting.
However, non-exclusive actions can be selected simultaneously. For
example, one can walk while eating.
Figure: Differentiating action selection from motor control in design.
It should be realized that, currently, motor control in complex animated creatures, such as articulated figures, still remains an unsolved problem. This limits the present feasibility of implementing the above described behavior control scheme in such creatures. However, this should not indicate the lack of generality of the control scheme for we believe that the separation of action selection from motor control is necessary in achieving high level behaviors in most creatures, especially complex ones.
| Xiaoyuan Tu | January 1996 |