Preface
With thirty-nine mechanical degrees of freedom the Proto-Android (PA) is at
the edge of effectiveness for a competent computer based system and aspires to
the lowest fringe of feeble primate capability. This report describes the
evolving rationale and implementation of PA's kinematics.
PA is a human sized ape-like platform designed for versatility, speed,
power, and economy. As example of PA's power and speed, its waist is composed of
three inch bore cylinders capable of kicking a thousand plus pounds at eye-blink
speeds. Hardware particulars such as actuators, sensors, and computational
resources involved in kinematic expression are found in related documents.
The performance bottleneck of PA-like robots is true intelligent control.
Industrial robots operate in highly structured environments and use canned
routines to complete adequate motions. More sophisticated methods using adaptive
control have gradually emerged from research addressing "unstructured"
environments such as planetary exploration. Techniques for trainable
self-optimization include neural nets, fuzzy logic, and every kind of
statistical and heuristic scheme.
These fancy tricks are desired approaches to Proto- Android's sensory and
kinematic behavior, but canned sequences are the initial goal and become a body
of Motion Language primitives described below. Adaptive responses are built on
these predefined elements. If decomposed to sufficient fine granularity, this
strongly constrained approach allows smooth behavior.
PA's Body:
PA's general success depends heavily on the judicious selection of its
geometric configuration. The major tradeoff brokered in PA's kinematic design is
between inherent stability and versatility. Stability reduces potential
mobility, but offloads the computation of balance, while versatility complicates
everything. The resulting compromise is a short legged, long armed beast that
generally walks on fours and is capable of planting on three limbs and
dexterously manipulating objects with the free right (fig.1). Limited
bi-pedalism is not excluded and will be attempted.
Many tradeoffs must be made to engineer a legitimate robot at this time. PA
is less flexible than real humans. Each DOF is limited to 60 degrees because of
engineering constraints like linearity of output. Cylinders add their limited
motions together to achieve required contortions.
The centered position all cylinders is a carefully chosen body
configuration to allow the greatest kinematic script space. It resembles the
dead man's float of a relaxing astronaut as identified by NASA ergometric
research. The intuition here is that the greatest functionality is preserved
when the centered position of all actuators corresponds to the ergometric
neutral configuration. This guess, like many others, takes the place of an
exhaustive and expensive inquiry for which resources are lacking.
Ralph Moser, the original master of modern robot mechanics, has personally
guided us in the pragmatics of developing something like PA. One of the author's
proudest, lucky, memories is when Ralph Moser came to the first Robofest.
Ralph was visiting his daughter here and saw the newspaper coverage for the
event and called to express the keenest, most vigorous interest in our modest
gig, even though he had dominated the field when most of us were not even a
gleam in anyone's eyes.
Word spread quickly in the Austin robotics community. In a flash a
symposium at UT was convened by Doc Robot (Dr. Del Tesar), at the UT Robotics
Lab. Ralph got the red carpet (tattered) at Robofest and paid us the highest
compliment by asserting that we had accomplished for "pennies" work on
a par with the best of the multi-million academic/military/industrial efforts in
the hairy field of bio-morphic robotics. This kindness was an inspiration to
continue to realize the compliment in projects like the ProtoAndroid.
The overall message that Ralph imparted is that while, in our times, we
cannot hope to duplicate nature's productions, we can with judicious reduction
to basic engineering practice, accomplish much.
PA's Hand:
A hand's complexity almost rivals that of the overall body. It is expedient
to design robotic hand systems simpler than human hands because most of the
functionality can be retained without all of the degrees of freedom of the real
thing. In the Robot Group we call this the "pinky factor". Design the
pinky in and you mostly get increased cost and complexity.
Pa's right hand design is intended to be roughly comparable to the MIT/Utah
hand. One less finger is the most obvious difference, in deference to
simplicity. HPL, mentioned above, and the general geometric conformance to its
model allows incorporation of preexisting development. Particular attention is
being paid to hand-eye research for wholesale adoption in which case the machine
vision must also be compatible. PA's left hand will be little more than a strong
claw on a 3 axis wrist. This handedness, again for simplicity, extends to sensor
resolution, and processing resources.
PA's right hand is a reductionist design with only two fingers and a thumb.
The advanced Salisbury/NASA hand uses this configuration to accomplish
impressive feats. Pa's hand does owe more to the MIT/Utah hand in terms of
documented research. PA's hand design has evolved to a bio-morphic shape that
resembles an advanced saurian manipulator. (see fig. 1)
PA's left hand is even more primitive, just a claw really, with no more
dextrous mission than assisting the right hand's actions. This is a further
savage savings of computational and mechanical complexity.
A particular structural challenge to PA's hand is that it has to serve also
as a "foot" during crawling. This crawling mode uses the heel of the
hand rather than the knuckle for structural reasons. Even so, hand walking is a
rough test for a manipulator and the hand is being built with carbon fiber,
kevlar, cables and laminated foams such as those used in running shoes.
PA's Head:
The head carries complex proximity sensors such as video cameras, sonar,
hearing etc. It rests on a standard 3 DOF joint built to 1/2 scale, (see report
1). In all likelihood PA will often have to freeze in order to take a steady "snapshot"
of the proximity as well as calibrate position information with the hand as
reference. A separate report on the multi-sensory suite goes into detail [].
Control:
The control paradigm for PA's hand derives from work with the MIT-Utah
dextrous hand done at Bell Labs []. A Hand Programming Language (HPL) was
developed whereby a minimal set of Motion Primitives are combined to form
complex sequential actions. This approach resolves the problem of dealing with
the trillions plus state spaces possible with a high DOF system. If one joint
has, say, 10 discrete positions, 10 joints combine to 10x10x10x10x10....
different configurations (excluding collisions). Consequently this scheme is
being applied to PA's entire body were body position primates combine to form a
Body Programming Language (BPL). HPL and BPL are to be augmented by other
techniques such as old fashioned robot arm "training", and
connectionist self optimization.
There are various classes of kinematic script. The simplest are canned
routines that range from the kinematic primitives from which all motions are
composed to elaborate sequences constituting gestures or default actions. As a
start PA will be able to crawl, backup, turn, gesture, grab, sit, or pick
himself up from a fall in a stereotyped fashion (fig.2). PA will generally
attempt manipulation or mapping by first planting his body in a stable three
point stance or a sitting position.
The most complex kinematic scripts involve self optimizing computationally
demanding interaction. Adaptive training requires the precautions of padded
surfaces, suspension lines, and strict safety procedures or skillful simulation.
The kinematic primitives underlying most motion scripts are the minimal
motions of which something meaningful or useful can be semantically defined. HPL
primitives like grab, push, and twist combine into scripts such as inserting a
light bulb. Each primitive is a tiny script of basic sensor/actuator operations.
In PA this approach is extended to the body in the form of BPL . It is in this
language that PA's canned locomotion, manipulation, and gesture are defined.
Other schemes may start from different premises that exist as options to BPL in
a given PA script.
A small number of kinematic heuristics are defined and act as a checklist
before scriptural expression. One is to avoid self and obstacle collisions,
another is to avoid extreme joint positions and a third is to detect falling and
respond by dropping its CG and extending a limb.
Another basic model for PA's default kinematics are the hopping,
somersaulting robots of Marc Raibert and colleagues at CMU. Their machines rely
on a real- time state machine representation that efficiently adjusts to small
variations in the task profile. PA will be larger, more loaded down, and more
sedate, at least at first but the cleanness and efficiency of the state machine
model will be exploited.
PA's basic locomotion modes are on two legs, if possible, and all fours,
come hell or high water. Locomotion and manipulation share resources. Locomotion
ends in a stereotypical tripod stance, freeing the dextrous hand for dedicated
use.
Sensory Feedback to Kinematic Operations:
Machine vision is impressive in manipulation tasks, especially when it
works at all in an unstructured environment. Pa will be rather blind at first
and kinematic performance will be correspondingly limited.
Attitude and acceleration sensing, contact detection and force feed back
are the primary sources of data for kinematic processes. Progress will be
incremental in squeezing performance from the sensor data. Current work involves
evaluation work by project engineers with two test-bed platforms.
Sensory feedback for kinematics is a real-time critical task. Much of the
sensory processing is therefore dedicated hardware and software, decoupling its
performance from the vagaries of the meta system. |
