Previous: Theoretical Issues Up: Related work
Brooks's subsumption architecture, [Brooks 1990][Brooks 1987][Brooks 1985], clusters behaviors into layers. Low-level behaviors, like deciding the direction of motion and speed, can be inhibited (subsumed) by behaviors determined at higher levels, such as avoiding obstacles. Subsumption behaviors are written as finite state machines augmented with timing elements. A compiler is used to simulate the operation of finite state machines and parallelism. This architecture is implemented on a variety of mobile robots. Frequently used behaviors are placed at a lower level than less frequently-used behaviors. This organization of behaviors gives the system fast response time and high reactivity. Our architecture is similar to Brooks's in our intra-level implementations of behaviors. However, the subsumption architecture lacks the separation we have made into conscious and non-conscious spheres. In anthropomorphic terms, Brooks's agents are all non-conscious. We believe that the off-line specification and compilation of behavior modules is too inflexible for autonomous agents that can adapt to a wide range of circumstances, especially if they have to learn from their interactions with the environment. Pattie Maes has experimented with a version of a behavior-based architecture, which she calls ANA [Maes 1991]. This architecture consists of competence modules for action and a belief set in a network relating modules through links denoting successors, predecessors, and conflicts. Competence modules have activation levels. Activations are propagated and the competence module with the highest activation level is given control. Maes has explored learning and has applied her architecture to robotic systems.
In the subsumption architecture, sensations and actions are abstracted by
giving them names like ``straightening behavior'' in order to make things
easier to understand for human observers. Much in the spirit of
[Agre 1988], we believe that behavior modules should more naturally
emerge from the interaction of the agent with its environment. In contrast
to hand coding behaviors and in order to facilitate embodiment, in GLAIR we
are experimenting with (unnamed) emergent behavior modules that are learned
by a robot from scratch. An (unnamed) behavior module can be thought of as
a set of tuples (P,A) where P is a set of grounded sensations and A is an
instance of an act. For instance, reaching for an object might be a set of
tuples (vision/sonar data, wheel motor actuation). After learning, this
new behavior module will become active only if the grounded sensations match
any of the grounded sensations experienced before. As a measure of
abstraction and generalization, we may allow near matches for sensations.
To bootstrap the learning process, we need a set of primary or first-order
(``innate'', for the philosophically inclined) sensations and actions. We
will return to this point briefly in section .
The Servo, Subsumption, Symbolic (SSS) architecture [Connell 1992] is a hybrid architecture for mobile robots that integrates the three independent layers of servo control, Brooksian behavior based modules, and a symbolic layer. Our architecture is similar to this in its general spirit of identification and integration of three distinct levels corresponding to levels of affinity-of-interaction (i.e., the rate at which it is in real-time contact with the world) with the outside world. This similarity also constitutes a point of departure, however, in that SSS is defined with respect to specific (and different) implementation techniques. For example, the symbolic layer in SSS seems to be a decision table versus a general KRR as intended in GLAIR. Unlike GLAIR, SSS assigns particular tasks for each layer and uses a hard-wired interconnection channel among layers.
Albus et al's hierarchical control architecture [Albus et al. 1981] is an example of a robotic architecture; we would say it is body centered. This architecture proposes abstraction levels for behavior generation, sensory processing, and world modeling. By descending down the hierarchy, tasks are decomposed into robot-motion primitives. This differs from our architecture, which is not strictly top-down controlled. Concurrently, at each level of the hierarchy, feedback processing modules extract the information needed for control decisions at that level from the sensory data stream and from the lower level control modules. Extracted environmental information is compared with the expected internal states to find differences. The differences are used for planning at higher levels.
Payton in [Payton 1986] introduced an architecture for controlling an
autonomous land vehicle. This architecture has four levels: mission
planning, map-based planning, local planning, and reflexive planning. All
levels operate in parallel. Higher levels are charged with tasks requiring
high assimilation and low immediacy. The lower levels operate on tasks
requiring high immediacy and low assimilation. Our architecture is similar
in this respect. The reflexive planning is designed to consist of pairs of
the form . Each
reflexive behavior has an associated priority, and a central blackboard
style manager arbitrates among the reflex behaviors. Some of the problems
with the earlier implementation due to using the blackboard model were
solved in [Rosenblatt \& Payton 1989].
Rosenschein and Kaelbling's work [Kaelbling \& Rosenschein 1990][Kaelbling 1988] describes tools (REX, GAPP, RULER) that, given task descriptions of the world, construct reactive control mechanisms termed situated automata. Their architecture consists of perception and action components. The robot's sensory input and its feedback are inputs to the perception component. The action component computes actions that suit the perceptual situation. We should note that unlike Brooks's behavior modules, situated automata use internal states, so their decisions are not Markovian (i.e., they are not ahistoric). They are mainly intended to produce circuits that operate in real-time, and some properties of their operation are provable. The mechanism for generating situated automata, although impressive, seems to inflexible for autonomous agents that have to operate in a wide variety of (possibly unknown) circumstances . Perhaps the operation of our Perceptuo-motor level could be modeled by a situated automaton, but we are not convinced that this is the right formalism to use, due to its inflexibility.
Gat in [Gat 1991] describes ATLANTIS, an architecture for the control of mobile robots. This architecture has three components: control, sequencing, and deliberation. The control layer is designed as a set of circuit-like functions using Gat's language for circuits, ALPHA. The sequencing is a variation of Jim Firby's RAP system [Firby 1987]. The deliberation layer is the least described layer. As with situated automata, we are not convinced that this is the right kind of formalism to use, for the same reasons.
An architecture for low-level and high-level reactivity is suggested in [Hexmoor 1989]. High-level reactivity is reactivity at the conceptual level. This architecture suggests that an autonomous agent maintains several different types of goals. High-level reactivity is charged with noticing impacts of events and actions in the environment on the agent's goals. Subsequently, high-level reactivity needs to guide the agent's low-level reactivity. Low-level reactivity is at the sensory, perceptual, and motor level. The mechanism for low-level reactivity is similar to other reactive systems that have components for perception and action arbitration. The novelty of this architecture is the incorporation of high-level reactivity and a supervisory level of planning and reasoning, which guides the choice of low-level reactive behaviors. In our present conception of agent architecture, we avoid a sharp separation between the two types of reactivity. We also relax the top-down nature of interaction between levels. Reactivity may be initiated at any level of our architecture either due to interaction with other levels or in direct response to external stimuli.
SOAR [Laird et al. 1987] was designed to be a general problem solving architecture. SOAR integrates a type of learning known as chunking in its production system. Recently, SOAR has been applied to robotic tasks [Laird et al. 1991]. In this framework, planning and acting is uniformly represented and controlled in SOAR. This approach lacks the ability of our architecture for generating behavior at non-conscious levels as well as the conscious level (or at different levels in general), and for having different-level behaviors interact in an asynchronous fashion. It also lacks our multi-level representations.
Simmons's Task Control Architecture (TCA) [Simmons 1990] interleaves planning and acting by adding delay-planning constraints to postpone refinement of planning until execution. For example, a plan for a robot to collect used cups for trash is decomposed into: navigate to the cup; pick it up; navigate to trash bin; deposit the cup. Since the robot does not have sensory information about the cup yet, the plan to pick it up is delayed until the robot gets close enough. Selectively delaying refinement of plans allows for reactivity. This type of ``stepwise refinement'' follows effortlessly from our architecture, without the need to explicitly implement it. Since conscious planning which goes on at the Knowledge level uses a more coarse-grained world model, there is simply no possibility to express fine details of planning and execution. These can only be represented and/or computed at the lower Perceptuo-Motor level and Sensori-Actuator level. Planning and execution in our architecture may proceed in a lock-step fashion, but they need not be. TCA uses a message-passing scheme among modules that allows concurrent execution of tasks. It has been used to control the six-legged walking robot Ambler and a cup-collecting robot.