Page 432 - Automotive Engineering Powertrain Chassis System and Vehicle Body

Decisional architecture C HAPTER 14.2

positions of the objects located in its environment. Then, planning phase, and consequently to be potentially able
the motion planner generated a collision-free path to cope with complex missions combining several goals
allowing the robot to reach the desired placement. and task constraints. As it will be shown in Section
Finally, the system controlled the robot actuators in order 14.2.2.3, this property does not hold when applying
to move it along the planned path, for a distance of about purely reactive approaches.
1 metre. Unfortunately, the complete SMPA process had
to be carried out after each motion of this type until the 14.2.2.3 Reactive architectures
goal had been reached, and each iteration required to
wait for about 15 minutes (mainly because of the image 14.2.2.3.1 The basic idea
processing time).
The main serious problems related to purely delib- Because of the above-mentioned strong limitations of the
erative architectures are the following: purely deliberative architectures, some researchers have
developed in the 1980s a new approach inspired by
The main drawback of this type of approach relies in ethology and entomology. Nowadays, we know that many
its intrinsic incapacity to cope with unpredicted live beings, in particular the insects, have very few ca-
events (mainly because of the large reaction time pacities for ‘modelling’ and ‘reasoning’. Despite this, they
which is required for processing the whole (SMPA) are able to achieve quite complex tasks; to an external
cycle). Consequently, it is almost impossible to take observer, they exhibit a global behaviour which seems to
into account dynamic objects or obstacles detected be the result of intelligence. On the basis of this obser-
while the robot is moving. The main reason for these vation, some researchers have proposed an alternative to
limitations comes from the slowness of the modelling deliberative architectures consisting in making use of
and planning phases, which cannot be carried out in purely ‘reactive behaviours’. This approach basically
real time, even when they are implemented on large consists in removing the ‘modelling’ and ‘planning’
computers external to the robot.
phases from the decisional loop, and in trying to produce
The second difficulty is related to the modelling ‘intelligent behaviours’ driven by sensory data: the robot
phase itself, which is in charge of reconstructing reacts intelligently to what it senses. Brooks (1990)
a model of the robot environment from sensory data. justifies the use of such an approach, by claiming that the
Indeed, this problem in its whole generality repre- ‘best model of the world is the world itself’.
sents a complete research domain which is still open The implementation of such an approach is based
(even if impressive results have already been upon the combination of several elementary modules
obtained by researchers in the held of computer implementing very simple (reactive) behaviours. Such
vision). It is well known, that relating sensory data to approaches are often referred as behavioural based
real objects is a difficult task because of the noisy, architectures: the observed behaviour of the robot is the
inaccurate and often spread nature of the informa- result of the combination of some various elementary
tion to process. This difficulty is related to the fact behaviours; it emerges from the interaction of the in-
that there is a great difference between sensing and volved elementary behaviours with themselves and with
perception. the environment. Each elementary behaviour (e.g.
The intrinsic differences which exist between the avoiding an obstacle or the heading to a goal) performs
model and the real world introduce strong uncer- a close coupling between the sensors and the effectors of
tainties on the positions/orientations of the robot and the robot. The intrinsic low complexity of the involved
of the obstacles. Taking into account these uncer- processing, along with the parallel structure of the be-
tainties is obviously necessary for obtaining a robust haviours, leads to a high speed execution property.
system. Unfortunately, this requirement makes the Note: The previous type of system is usually referred
planning phase much more complex, and poses sev- as a ‘reactive control system’. However, one can find two
eral modelling and algorithmic problems which are different definitions of reactivity in the literature: for
still open. peoples in the fields of computer science and robotics,
Consequently, the use of such an approach seems to be a reactive system is ‘a system able to react continuously
limited to the case of a robot evolving in a static, strongly to a physical environment, at a speed determined by this
constrained and a priori known environment. This is why environment’ (Harel and Pnueli, 1985); for peoples in
the purely deliberative approach is not used anymore in the field of cognitive sciences, an agent is said to be re-
recent robot control architectures (even if it has raised active ‘if it does not have an explicit representation of the
key research issues that are still actively studied). world’ (Ferber, 1995). Even if they look different, these
However, it should be stressed that the deliberative two definitions are not contradictory, since they imply
approach has a significant advantage: it includes the that the involved controllers react directly to the stimuli
property to apply high-level reasoning capabilities at the coming from the physical world; hence avoiding the

439