Page 435 - Automotive Engineering Powertrain Chassis System and Vehicle Body

CHAP TER 1 4. 2 Decisional architecture

execution monitoring parameters. This motion plan is However, only the implementation of some expert
obtained by simulating the displacement of the robot agents belonging to the ‘Reflexive planner’ (e.g. follow
along the planned collision-free path, and by anticipating a wall, or avoid an obstacle) has been deeply described. In
the possible failures and the sensory data to monitor. The this implementation, the related reflex behaviours are
parameterized motion plan is finally used by the System associated with some virtual sensors in charge of pro-
Executive to monitor the execution of the robot task. In viding a specialized information (e.g. obstacle detection,
practice, only some predefined and simple ‘Reflex object recognition, localization). Then, the activation of
Actions’ can be introduced into the motion plan (e.g. the appropriate reflex behaviours is done using a black-
stopping the robot and moving back to a safe position). board technique and some predefined priorities.
Such an architecture basically applies the sequential Quite complex missions have been planned and exe-
SMPA paradigm, while executing some predefined reflex cuted with a significant level of reactivity using this
manoeuvres when dangerous situations have been approach. The main limitations of the system come from
detected. This approach represents to some extent the both the limited communication mechanism existing
minimum level of integration of a reactive component between the different layers, and the predefined com-
into a deliberative architecture. bination of behaviours. Later on, Payton et al. (1990) and
Payton’s architecture This architecture (Payton, 1986) Rosenblatt (1997) have improved the reactivity of the
is based on a hierarchical decomposition, in which each system by using a distributed control arbitration tech-
layer is characterized by a type of sensory data processing nique, allowing them to combine controls coming from
(modelling). As shown in Fig. 14.2-3, this architecture is both the reactive behaviours and the planning layers.
composed of a layered perception system and of four Task control architecture (TCA; Simmons) The TCA
main decisional modules: (1) the ‘Mission planner’ de- proposed by Simmons (1994) represents a new alterna-
fines a sequence of geographical goals to reach along with tive to traditional hierarchical approaches. This architec-
their associated motion constraints; (2) the ‘Map-based ture is composed of an arbitrary number of specialized
planner’ uses the global world model to generate paths modules, communicating through messages with a central
connecting the previous geographical goals (the response management module. The specialized modules carry out
time of this planner is of a few minutes); (3) the ‘Local the tasks which are specific to the robot to control,
planner’ determines the details of the motions which are whereas the central management module supervises the
required for moving the robot along the planned paths functioning of the whole system and controls the routeing
(the response time of this planner is of a few seconds); of the messages between the various modules; messages
(4) the ‘Reflexive planner’ controls in real time the ex- can be used for an information request, for sending
ecution of the motion task. a command, or for asking for a task decomposition to the
From the implementation point of view, this archi- planners. The TCA architecture makes use of a hierar-
tecture has been developed using expert agents com- chical representation of the tasks/sub-tasks relationships
municating between them using a blackboard technique. (called the ‘task tree’) for maintaining an internal repre-
Using this approach, the activity of a particular module sentation of the robot task to execute.
can theoretically be controlled by a higher layer through Fig. 14.2-4 shows how the TCA architecture has been
the selection of the expert agents to be activated. implemented for controlling the Ambler legged robot.
However, this implementation put the emphasis onto the
planning functions (gait planner, footfall planner, etc.), and
reduces the reactivity to the processing of some excep-
tions (for stabilizing the robot). The main drawback of this
architecture relies on the centralized processing schema
and its associated communication mechanism, which
often implies rather long response times incompatible
with fast robots. This is why the author has also imple-
mented additional (i.e. apart from the TCA architecture)
some ‘emergency reflexes’ for quickly stabilizing the
Ambler robot when a problem arose.

14.2.2.4.3 Reactive-based hybrid architectures
AuRA architecture (Arkin) The AuRA architecture pro-
posed by Arkin (1987; 1989; 1990) is mainly based on
the concepts of ‘motor schema’ and ‘perceptive schema’,
Fig. 14.2-3 Payton’s architecture. which are used for describing the links existing between

442