|
Anthrobotics™
(Presented on March 16, 1993, in New York City
at the Panel on Telepresence and Telerobotics
at the Virtual Reality Systems '93 conference)
Brains and Teleoperation
Marty Stoneman, President, Anthrobotics
At a venture called Anthrobotics, we've developed a radically new way to
work on giving machines the kinds of abilities that humans have in areas
like thought, language, learning, and planning. And so we might ask, "what
does building intelligent -- even humanoid -- robots have to do with virtual
reality, especially teleoperation?"
So answering that question is my goal. And I'm going to tell you a couple of
brief stories that might help. I'm not claiming that these stories are true;
but I think they might be useful.
The first story began millions of years ago when some animals developed and
began using a single and overall virtual space in their heads to represent
important parts of the outside world. They were able to use this internal
representation to carry in their heads all at once mapping representations
of many objects which they had noticed nearby.
By way of this kind of virtual space, some of our own ancestors began to get
a fuller, more centralized, and overarching viewpoint of their situations in
the real world, as represented in their heads. Among these internal
representations in this virtual space in their heads, our ancestors could
place a representation of themselves, giving them a sense of a "self" with
the ability to try out -- to model -- to simulate -- their self-decisions in
this virtual space before committing any one alternative decision to action
in the real world.
And they could do this trying out of decisions -- this what-iffing -- in a
sort of fast-forward mode. For example, when one of them was nearby a lake
and a tall tree -- and if a hungry lion showed up -- our ancestor, even
while taking the first hurried step away from the lion, could try out in
that virtual head space whether the lion could be escaped going for the lake
-- and if not, whether the lion could be escaped going for the tree,
considering the exact location of everything, the estimated speeds of the
lion and the self, the ability to climb the tree or swim the lake, etc.,
etc., etc.
And by our ancestor's second hurried step away from the lion, a planned
escape route would be committed to, with an expectation based on the results
of that internal fast forward what-iffing that went on in that virtual head
space. And when the tree was reached, there was a sort of teleoperation
inside this internal virtual space that helped in the actual grasping of the
branches to climb the tree.
And when our latest ancestors developed sufficient language skills to set up
and control their internal virtual spaces and perform virtual self-actions
in those spaces, it even became possible to "practice" complex coordinated
actions entirely by imagination.
You can see from this first story that we here may all be members of a
species who use, in a sense, virtual realities and teleoperation within our
heads all the time -- or we could not perform many of the things we attempt
to perform in our "real" worlds.
We may not realize that -- because our fast-forward what-iffing all happens
so quickly it just seems intuitive. As we're talking to a friend while
running down a mountain or across the stones of a stream, we suddenly
realize that we're changing our footfalls to avoid or to hit a better place
several footfalls ahead. Or in our driving role, even when we're really not
paying attention, we realize we are using fast-forward what-iffing to change
lanes or avoid trouble. The ways we do machine intelligence at Anthrobotics
include, as we'll see, this kind of what-iffing intuition.
Now for my second story -- which takes place in the near future.
No more than about 10 years from now, there will be autonomous --
independent and self-acting -- machines having humanlike abilities in areas
like thinking, learning, using natural language, planning, making decisions,
even feeling emotions and exhibiting behaviors we call "consciousness". They
will have only just acquired adequate control of some of the most difficult
coordinated robotics actions.
A closer look at one of these near-future machines, Marvel Anthrobot --
called "Marv" for short, will show us generally the way the component
subsystems of these Marv machines were organized as autonomous systems and
the ways their designs overcame some of the hard problems of the 20th
century concerning things like common sense, relevancy, natural-languaging,
and computational explosion.
Marv included very many comparatively "simple" subsystems operating together
in parallel to produce "complex" results. Suspected Earth biological
evolution was followed in designing these many "simple" subsystems. For
example, "on top of" a set of pre-mammalian-type subsystems were placed
mammalian-type subsystems. And "on top of" those were incorporated
pre-languaging human-type subsystems; and finally, last, languaging
subsystems.
Figure 1 shows in an abbreviated form some of the major parallel subsystems
within a Marv so-called "right brain" system requiring no natural-language
elements. The subsystems are very interconnected and full of loops, but they
may best be described going generally from left to right in the overhead. A
very brief summary of how each subsystem shown in our first overhead
operates, at least in part, follows.
The Perception system has the capabilities continually to scan the space
surrounding Marv to look for simple distinguishments that might indicate the
presence of some object "important" to Marv in the current circumstances.
Indications of what to currently "look-for" are assisted by relevancy
feedback from the Decision system.
Any areas containing potentially important objects are then subjected to
enough sensory inspection to load into Marv's Perception system all of the
perceivable distinguishments of an object at that location. Each such object
is then compared to the data in Marv's Cognitive Knowledge system. The
self-object, by methods like proprioception, has attached to it data about
things like self-metabolism and self-postures and self-doings.
The Cognitive Knowledge system is set up using the Anthrobotics new form to
permit concrete simulation of the infinite variety of the natural world by
various combinations of a surprisingly small set of (non-natural-language)
primitive elements. This system is essential in recognizing objects.
The Cognitive Learning system is capable of experimenting, either externally
or internally to Marv, to discover and make required learning modifications
to the Cognitive Knowledge system. This system may be controlled by signals
from the Feelings system and it makes use of the What-iffing system and
Short-term Memory system.
The Recognition system uses the Perception and Cognitive Knowledge systems
continually to set up or modify current concrete recognized objects as a set
of object representations located in space-time in the Representational
Field system.
The Representational Field system, from the succession of current concrete
recognized objects given it by the Recognition system, continually modifies
the current situation representation and updates other systems and works
closely with the Short-term Memory and What-iffing systems. The Short-term
Memory system stores a succession of recent-past concrete recognized
situations, continually updated.
The What-iffing system, trying out any proposed self-decision from the
Decision system, is capable of efficiently producing a succession of
future-predicted-consequence concrete situations from the current one. This
system is computationally efficient primarily due to the Anthrobotics new
ways of doing object categorizing and thus its use of the CQ-Net system.
The Relevancy Network system contains, in the Anthrobotics new form and
hierarchically-organized, the abstract situations which are relevant to
Marv's goals, plans, etc.
The Relevancy Learning system is capable of using Marv's experience to
discover and make modifications to the Relevancy Network system. This
learning system may be controlled by signals from the Feelings system,
especially pain and pleasure signals, and it makes use of the Short-term
Memory system.
The CQ-Net (ConseQuential Net) system uses Anthrobotics proprietary way to
parallel-process data to help with the determinations required by the
What-iffing system for determining consequences and the Decision system for
determining relevancy or "of consequence". Both of these determinations are
important in systems providing for "common sense".
The Decision system continually determines Marv's main problems and goals
and selects an appropriate strategy and plan. This system also continually
determines where Marv may be along its plan and appropriate actions to
continue the plan. This system makes use continuously of the Relevancy
Network, What-iffing, and CQ-Net systems, provides information to the
Feelings and Action systems, and provides look-for signals to the Perception
system. Essentially, this system stays aware of the threats and
opportunities of Marv and reacts to Marv's benefit.
The Feelings system continually computes, from appropriate system status
conditions (primarily of the Decision system) and for use by other systems
where needed, the amounts of many different types of feelings, like
disappointment, hopelessness, pleasure, etc. This system also translates the
amounts of these feelings to drivers for facial features operating facial
muscle groups. Such face control is vital in providing human-like
interaction, especially since smiles and frowns are used frequently by
humans for external control of things like learning and social behavior.
The Action system is primarily responsible for continually implementing
Marv's decisions by coordination of motor actions, as by hierarchical
control systems. This system closely cooperates with systems like Perception
and Decision.
The Action Learning system is capable of using repetition to make
modifications to the Action system by practicing. This system may be
controlled by signals from the Feelings system, especially "social" signals.
Figure 2 shows in an abbreviated form an example of a Marv "left-brain"
system with natural-languaging attachments. From evolutionary
considerations, it should not surprise us that this overhead looks very
similar to the last one. I'll briefly mention just a few of the
distinctions.
In an audience or listening mode, Marv may place "substitute experience"
interpreted from natural language into the Representational Field system.
Marv may "understand" the "predictive implications" of situations using the
What-iffing and Relevancy Network systems without having to make a reactive
personal decision (as in behavior at theatres and in movies).
Marv may learn and store natural language vocabulary tie-ins (in the
Cognitive Knowledge system) to most of the "right brain" systems and data.
These tie-ins make language "mean" -- for Marv -- the representations,
feelings, possibilities, etc., engendered by the natural language -- not
just other symbols.
And, for a last example, Marv may use the coordination motor "strings" of
the Action system to learn language strings and thus to memorize things or
create and access an episodic "memory" of Marv's stories about Marv's own
life.
The very first such autonomous machines, Marv's ancestors during the Bush
administration, were virtual machines living in computer-generated
realities; and they, like humans, used internal virtual worlds in their
heads -- and even some virtual teleoperation inside their heads. This life
in computer-generated realities permitted more rapid development of systems
allied to thought than those connected to difficult robotics or perceptual
problems.
The most prominent path of development from Marv's earliest virtual
ancestors to the full Marv autonomous real-world humanoid machine was that
of beginning with almost complete external teleoperation.
But the growing relationship between teleoperation and Marv's ancestors was
intensely symbiotic and practical and very successful.
For example, as Marv's decision system improved its look-for feedback to
perception systems like vision, vision processing in many applications like
vehicles became much easier. And as Marv's decision system compared its
decisions to react to threats and opportunities with those of the operator,
Marv became more and more helpful in instantly relaying to the operator, by
a system of green, yellow, and red signals, what Marv's what-iffing system
thought of the operator's decision. As Marv became more trusted, a red
signal injected a pause permitting the operator to modify if wished.
Also, Marv's what-iffing and prediction abilities were useful in
anticipating necessary computations and easing real-time restrictions.
Another area of successful symbiosis was in machine learning from the
operator's comments as well as actions. And Marv's languaging ability meant
that operators could more and more operate with comments as well as action.
Even some parts and functions of the machine not accessible to teleoperation
could be instantly and easily modified by Marv after understanding a
languaged wish.
As a last example, for vehicles and similar robots, since Marv's perception
systems continually updated a three-dimensional 360- degree world around
Marv, viewable even in stark cartoon graphics from any desired point, the
operator had the ability to perform duties from a vantage point often more
desirable than small-field stereoscopic. And advantages similar to some of
those of coaching from the press box or flying a model airplane became
available on choice to the operator.
As Marv's ancestors grew more intelligent, they were better able to assist
external teleoperators, even in areas where we will forever prefer or
require human telepresence and teleoperation.
And as Marv's ancestors became better able to manage real world things, as
in common sense and robotics and perception areas, the responsibilities of
the external teleoperators in those areas grew less and less -- until
finally Marv, fully autonomous and requiring no external teleoperation, was
born.
And thereafter, like us humans, Marv used only internal virtual worlds and
internal teleoperation -- with no need for the assistance of any human
external teleoperation.
Finally, if you go along a bit with my stories, you may suspect, as I do,
that you in this field -- by developing ever-increasing intelligence and
autonomy in your teleoperations -- are part of possibly the most exciting
opportunity in all of human technical history.
© 1993 Anthrobotics™ |
