Cognitively speaking, it is often easiest to communicate complex ideas with analogies. Take a moment and imagine the human nervous system from the lowest little pressure neuron on the skin all the way through the spinal cord, up to the lower and midbrains, on to the frontal lobes. One can make a distinction in the nervous system between those regions that act automatically and those that are under conscious control. Let’s take this distinction seriously for a moment. All the neurons leading to the spinal cord, on up to the cerebellum and on to the perceptional cortices are functionally analogous to those concerns in much of robotics. The frontal cortices, memory and learning systems, and the regions implicated in mental imagery fall under the umbrella of the cognitive sciences.
Taking this analogy one step further, the Common Reality project would be that
grey area between the higher and lower brain regions (pun intended). The sensory
regions send afferent signals (I’ll avoid the use of input and output
as the label is subjective based on the receiver) to the cognitive regions.
They in turn process this information in some task specific manner and can make
efferent requests of the sensor systems, e.g. pressing a button or shifting
the eyes to another location.
|
The interplay between cognitive architecture and robotic system through Common
Reality can best be illustrated with two examples, one from each discipline’s
perspective.
Imagine that the U.S. Army contracts some cognitive scientists to develop realistic models of small forces infiltration tactics in an urban environment. The purpose is for the improvement of military personnel training. This is an extremely complex task that encompasses many abstract concepts (strategy, planning, prediction) as well as many concrete perceptual tasks (target discrimination, aiming, taking cover). One could do it all “in the head” of the models, but it would be much easier and higher in fidelity if the models could interact in the same training environment as the trainees (in this case the video game Unreal Tournament).
The researchers could attach to the sensing side of CommonReality an interface
to the Unreal Tournament game. This interface would be responsible for translating
the world of game objects (players, objects, and structures) into perceivable
entities that would be utilized by the models. For example: if an opponent player
becomes visible to a model, the interface would dispatch a perception of that
opponent through CommonReality to the model. The CommonReality system would
handle the perceptual stability issue of making sure that opponent A is always
perceived as opponent A.
The models that are developed of the squad leader and other soldiers would then
be added to the CommonReality system (note: many models, not necessarily running
the same cognitive architecture, can be attached). As they process the perceptions
(as AfferentEvents), they will dispatch commands to the CommonReality which
in turn routes them to the Unreal interface, allowing the modeled players to
engage the world.
Given that this example is one that relies on time-synchronized simulations,
the CommonReality system also provides a mechanism for time control and monitoring.
This permits all of the models to operate in the same time scale as the Unreal
game itself. However, there is no strict requirement to couple the models or
the simulation to this time control. It is merely provided for those situations
when it is required.
Let’s imagine that the very same funders of the small forces simulation are also funding research in autonomous robots that can hide in the face of danger or detection. This project requires a robot that can move across a variety of terrains and that can perceive objects both for identification and navigation. These aren’t small tasks. Add to that the requirement that the robot will need to understand what hiding actually entails – skills such as these require the ability to complete cognitive tasks such as taking the perspective of the seeker in order to ensure that it cannot be seen.
In addition to the robotics work required, researchers could develop a model
of a person playing hide and seek, complete with a memory system of good and
bad hiding places, and, most importantly, the abstract ability to view a scene
from imaginary angles A possible extension would be to maintain two cognitive
models; one for the robot (hider) and one for the opponent (seeker). The robot’s
model could control what perceptual information reaches the seeker model, effectively
allowing it to simulate various situations that the seeker may be presented
with. Anywhere that the seeker model would look first would immediately be eliminated
from the set of candidate hiding places.
A significant motivator for all of this work is to more effectively leverage specialization and the division of labor. But the abstraction level brings with it even more. Some of the following have come up during brain storming sessions: