Subsumption Architecture

Subsumption Architecture

Rodney Brooks’ subsumption architecture is the most influential of the purely
Reactive Paradigm systems. Part of the influence stems from the publicity
surrounding the very naturalistic robots built with subsumption. As seen
in Fig. 4.5, these robots actually looked like shoe-box sized insects, with
six legs and antennae. In many implementations, the behaviors are embedded
directly in the hardware or on small micro-processors, allowing the
robots to have all on-board computing (this was unheard of in the processorimpoverished
mid-1980’s). Furthermore, the robots were the first to be able

to walk, avoid collisions, and climb over obstacles without the “move-thinkmove-
think” pauses of Shakey.
The term “behavior” in the subsumption architecture has a less precise
meaning than in other architectures. A behavior is a network of sensing and
acting modules which accomplish a task. The modules are augmented finite
state machines AFSM, or finite state machines which have registers, timers,
and other enhancements to permit them to be interfacedwith othermodules.
An AFSM is equivalent to the interface between the schemas and the coordinated
control strategy in a behavioral schema. In terms of schema theory,
a subsumption behavior is actually a collection of one or more schemas into
an abstract behavior.
Behaviors are released in a stimulus-response way, without an external
program explicitly coordinating and controlling them. Four interesting aspects
of subsumption in terms of releasing and control are:
1. Modules are grouped into layers LAYERS OF of competence. The layers reflect a hi-
COMPETENCE erarchy of intelligence, or competence. Lower layers encapsulate basic
survival functions such as avoiding collisions, while higher levels create

more goal-directed actions such as mapping. Each of the layers can be
viewed as an abstract behavior for a particular task.
2. Modules in a higher LAYERS CAN SUBSUME layer can override, or subsume, the output from be-
LOWER LAYERS haviors in the next lower layer. The behavioral layers operate concurrently
and independently, so there needs to be a mechanism to handle
potential conflicts. The solution in subsumption is a type of winner-takeall,
where the winner is always the higher layer.
NO INTERNAL STATE 3. The use of internal state is avoided. Internal state in this case means any
type of local, persistent representation which represents the state of the
world, or a model. Because the robot is a situated agent, most of its information
should come directly from the world. If the robot depends on
an internal representation, what it believes may begin to dangerously diverge
from reality. Some internal state is needed for releasing behaviors
like being scared or hungry, but good behavioral designs minimize this.
4. A task is accomplished by activating the appropriate layer, which then
activates the lower layers below it, and so on. However, in practice, subTASKABLE
sumption style systems are not easily taskable, that is, they can’t be ordered
to do another task without being reprogrammed.