Previous: The Knowledge Level Up: Architecture levels
Next: The Sensori-Actuator Level
The Perceptuo-Motor level uses a more fine-grained representation of events, objects, and states of affairs. For instance, they specify such things as size, weight, and location of objects on the kinematic side, and shape, texture, color, distance, pitch, loudness, smell, taste, weight, and tactile features on the perceptual side. At this level, enough detail must be provided to enable the precise control of actuators, and sensors or motor memory must be able to provide some or all of this detail for particular objects and situations. The Perceptuo-Motor level is partly aligned with the Knowledge level, in that there is a correspondence between object identifiers at the Knowledge level and objects at the Perceptuo-motor level.
Kinematic and perceptual representations of particular objects or typical object class instances may be unified or separate, and both kinds of representations may be incomplete. Also at this level are elementary categorial representations; the kinds of representations that function as the grounding for elementary grounded symbols at the Knowledge level, i.e., sensory-invariant representations constructed from sensory data by the perceptual processor [Harnad 1990].
The representations at this level are embodied (cf.
[Lakoff 1987]), meaning that they depend on the body of
the agent, its particular dimensions and characteristics. Robots will
therefore have different representations at this level than people would,
and different robots will have different representations as well. These
representations are agent-centered and agent-specific. For instance, they
would not be in terms of grams and meters, but in terms of how much torque
to apply to an object to lift it, or what
percentage of the maximum to open the hand to grasp an object. Weights of
things in this kind of representation are relative to the agent's lifting
capacity, which is effectively the maximum weight representable. An agent
may have a conscious (Knowledge level) understanding and representation of
weights far exceeding its own lifting capacity, but that is irrelevant to
the Perceptuo-Motor level. When it comes to lifting it, a thousand-pound
object is as heavy as a ten-thousand-pound one, if the capacity is only a
hundred or so. Similarly, sizes are relative to the agent's own size.
Manipulating small things is not the same as manipulating large things,
even if they are just scaled versions of each other. A consequence of using
embodied representations is that using different ``body parts'' (actuators
or sensors) requires different representations to be programmed or
(preferably) learned. While that may be a drawback at first, once the
representations are learned they make for faster processing and reactive
potential. Representations are direct; there is no need to convert from an
object-centered model to agent-centered specifications. This makes the
computations at this level more like table lookup than like traditional
kinematics computations, which can be quite involved. Learning new
representations for new objects is also much simpler; it is almost as easy
as trying to grasp or manipulate an object, and merely recording one's
efforts in one's own terms. The same holds, mutatis mutandis, for
perceptual representations.
There are a number of behaviors that originate at this level: some are performed in service of other levels (particularly deliberative behaviors), some are performed in service of other behaviors at this level, a few are ongoing, and some others yet are in direct response to external stimuli. An agent may consciously decide to perform Perceptuo-motor actions such as looking, as in look for all red objects, or to perform a motor action, such as grasp a cup. These actions originate at the Knowledge level and are propagated to this level for realization. An agent has to perform special perceptual tasks to serve other behaviors, such as to find the grasp point of a cup in order to grasp a cup. These perceptual tasks may originate at this or another level.
At the Perceptuo-Motor level, an agent has a close coupling between its behaviors, i.e., responses, and stimuli, i.e., significant world states. We observe that, for a typical agent, there are a finite (manageably small) number of primitive (``innate'') behaviors available. As the agent interacts with its environment, it may learn sophisticated ways of combining its behaviors and add these to its repertoire of primitive behaviors. We will consider only an agent's primitive abilities for now. We further assume that the agent starts out with a finite number of ways of connecting world states to behaviors, i.e., reflex/reactive rules. Following these observations, we suggest that at this level, the agent's behavior-generating mechanism is much like a finite state automaton. As we noted earlier, learning will change this automaton. The agent starts with an automaton with limited acuity, and uses its conscious level to deal with world states not recognizable at the Perceptuo-Motor level. For instance, the Perceptuo-Motor level of a person beginning to learn how to drive, is not sophisticated enough to respond to driving conditions automatically. As the agent becomes a better driver, the conscious level is freed to attend to other things while driving. This is called automaticity in psychology. We discuss an implementation mechanism for these automated behaviors later in this paper.