Page 433 - Automotive Engineering Powertrain Chassis System and Vehicle Body

CHAP TER 1 4. 2 Decisional architecture

construction of a ‘non-correct’ model of the world, and structure which potentially allows the implementation of
enabling high execution speeds. The next subsections new higher levels of competence and (4) a simple imple-
illustrate some typical implementations of this type of mentation which drastically reduces the production cost
architecture. and makes miniaturization possible (see for instance the
proposal of Brooks for exploration robots (Brooks and
14.2.2.3.2 Subsumption architecture Flynn, 1989)).
However, the experience has proven that such an
R. Brooks is one of the pioneers in the domain of reactive
architecture can hardly be fully implemented (except
architectures for robots, and his work on this topic is
the lower levels of competence), and consequently it
certainly the most known in our scientific community
(Brooks, 1990). His behavioural architecture is based on exhibits strong limitations coming from its lack of high-
a vertical decomposition into several levels of competence level reasoning. Indeed, even if this capacity is theoreti-
(see Fig. 14.2-1). Each of these levels represents an in- cally present in the highest levels of competence, these
dependent behaviour, receiving data from sensors and levels have never been implemented and the feasibility
acting on the robot actuators. The first implementation still remains doubtful. Consequently, only simple behav-
of this architecture has been realized using augmented iours can be implemented in the current state of the art.
finite state machine (AFSM) models, a communication
mechanism based on messages, and an inhibition/sup- 14.2.2.3.3 Other reactive architectures
pression mechanism for modifying dynamically the The supposed potentialities of the behavioural approach
stimulus or response signals of some active modules. have motivated various developments and theories.
It should be noticed that such an architecture does not As has been previously mentioned, the subsumption
have any central decisional kernel for selecting the architecture uses a hierarchy of behaviours, in which only
required behaviours: this choice is continuously done at one module at a time ‘controls’ the robot (by possibly
execution time using the structure of the implemented integrating the results of some other behaviours located
layers and communication network. lower in the hierarchy). Another possible approach, as
Eight levels of competence have been defined by the one proposed by Anderson and Donath (1990),
Brooks as a guide for his work (see Brooks (1990) for more consists in combining the controls simultaneously
details): the lower levels of competence consist in (1) ‘recommended’ by several behaviours. In such an ap-
‘avoiding obstacles’, (2) ‘wandering aimlessly around’ and proach, each reactive module implements a simple
(3) ‘exploring the world’; the higher levels consist in (7) ‘reflex behaviour’ having no memory capacity. Such
‘formulating and executing plans’, and (8) ‘reasoning a type of behaviour has been defined by McFarland
about the behaviour of objects and modifying plans ac- (1987): ‘a reflex behaviour is the simplest form of re-
cordingly’. This approach has been used to develop several action to an external stimulation; stimuli such as
small autonomous robots at MIT. However, only the first a sudden change in the level of illumination or a contact
three levels of competence have been implemented. arising at a given point of the body, generate an
Such an architecture exhibit some interesting proper- automatic, involuntary and stereotyped response.’ In the
ties underlined by Brooks: (1) a short response time which approach of Anderson and Donath, the involved modules
provides the robot with the capacity to move in a dynamic generate artificial potential fields whose composition is
environment, (2) a robust controller which can potentially used to determine the motion direction to be followed by
work even if one of the modules does not work correctly the robot (by using the classical gradient technique). The
(this is due to the parallel structure and the relative set of the behaviours to be considered at a given time is
independence of the behaviours), (3) an incremental selected according to criteria taking into the account the

Fig. 14.2-1 Principle of the subsumption architecture (Brooks).

440