Page 434 - Automotive Engineering Powertrain Chassis System and Vehicle Body

Decisional architecture C HAPTER 14.2

are referred to as hybrid architectures. However, several
types of hybrid architecture can be distinguished
depending on the way the deliberative and reactive
components have been combined. A commonly used
classification (with some minor variations) in the litera-
ture consists in considering three main types of
approaches:
Deliberative-based hybrid approaches. In such app-
roaches, the planning function has a predominant
role (i.e. motion is driven by planning using the
SMPA paradigm), and the reactive functions are only
added for dealing with some exceptions. The related
elementary reactive actions can either be integrated
as part of the architecture, or inserted into the gen-
erated motion plans by the planner.
Reactive-based hybrid approaches. In this type of
approach, the robot motions are executed under
the supervision of the reactive functions, and the
planning functions are mainly used as a ‘guidance
Fig. 14.2-2 JPL exploratory robot architecture (Gat).
resource’ for the reactive component of the system.
Three-layered hybrid approaches. Most of the current
task to be achieved and the characteristics of the envi- developments on robot control architectures are
ronment. In a similar way, Rosenblatt and Payton (1989) based on this type of approach. Such an approach
have proposed to use an artificial neural network to may be seen as the implementation of an ‘adaptable
combine the controls recommended by the selected planning-reacting scheme’. The basic idea consists in
behaviours; however, their system has been developed adding an intermediate layer for appropriately
using an empirical approach, and the way the controls interfacing the deliberative and reactive functions.
have to be appropriately combined is rather difficult to Then, one can consider that this type of architecture
define. relies on the three following basic logical functions:
In order to overcome the previous difficulty, some planning, sequencing and reacting. In such a para-
authors have tried to use training techniques for realizing digm, the behaviours of the reactive layer are con-
the selection mechanism. For instance, Humphrys ditionally instantiated by the sequencing module,
(1995) has proposed to apply a learning phase to each according to some sensing conditions and constraints
individual behavioural module, in order to construct defined by the planning module.
a function for evaluating the ‘quality’ of the proposed
actions; then this function is used on-line to select the
best ranked behaviours. Lin (1993) developed a similar 14.2.2.4.2 Deliberative-based hybrid
approach by using Q-learning techniques for imple- architectures
menting the training phase. JPL exploratory robot architecture (Gat) This architecture
Although some improvements have been proposed for has been proposed by Gat et al. (1990) at Jet Propulsion
the behaviours selection and composition mechanisms, Laboratory (JPL), for providing a planet explo-
all these approaches exhibit the same general limitations ratorymobilerobotwithsomepartialautonomycapabilities
as the subsumption architecture (see Section 14.2.2.2).
(since such a robot cannot be directly teleoperated from
the earth because of communication delays). As shown in
Fig. 14.2-2, this architecture is based on four main
14.2.2.4 Hybrid architectures modules: the Perception Module, the Path Planner,
the Execution Monitoring Planner and the System
14.2.2.4.1 How to hybridize? Executive.
The purely reactive and purely deliberative approaches The main task of the Perception Module is to build
represent two extremes that many authors naturally a local map of the environment, using sensory data and
tried to combine. The objective is to try to preserve the global data provided by the orbiter. Then, this map is
potential high-level reasoning capacity of the deliberative used by the Path Planner and by the Execution Moni-
approaches, while ensuring the robustness and the short toring Planner to respectively plan a collision-free path
response time of reactive approaches. Such approaches (of roughly 10 metres) and a complete plan including

441