docs/NME3-1Final.pdf. It's a snapshot of the page taken as our search engine crawled the Web.
The web site itself may have changed. You can check the current page or check for previous versions at the Internet Archive. Yahoo! is not affiliated with the authors of this page or responsible for its content.
A neuromuscular controller for fast, dynamic bipedal walking
V o l u m e  3  
 N u m b e r  1  
M a y  2 0 0 6
A neuromuscular
controller for fast,
dynamic bipedal walking
In this issue:
A neurobiological perspective
on intelligent devices
Freeing vision from frames
Light touch for balance
AER representation tools
Laboratory Notes
Unconventional functions in
charge-based CMOS

Workshop Review
INE/UPenn word-serial AER
workshop
Although biped robots have been devel-
oped using various technologies, they are
still well outperformed in some important
aspectssuch as speed and robustnessby
their natural counterparts, humans. In hu-
man and animal walking control, stable
gaits emerge from the global entrainment
between the neuro-musculo-skeletal system
and the environment.
1
Moreover, in human
and animal locomotion, the muscle is more
than a simple actuator. Some special prop-
erties of musclesthe inertia of a limb or
the elasticity of a muscle, for examplecan
tremendously simplify the control demands
of the nervous system for walking.
Here, we present our design of, and
experiments with, a planar biped robot and
its reflexive neuromuscular control network.
The network is composed of biologi-
cally plausible model neurons and a simple
muscle model that is simulated with a con-
trol algorithm implemented on DC-geared
motors. In contrast to other walking robots,
our design has no
central pattern gen-
erator in the form of
a neuronal oscillator.
Rhythmic patterns
are generated by the
whole system using
the electrical and me-
chanical properties
of the motors, the
limbs, and the envi-
ronment. In the ex-
periments, our biped
robot attained a rela-
tive walking speed
faster than any other
current biped robot,
and comparable to that of humans.
The robot design
RunBot is 23cm high, foot to hip-joint
axis. It has four joints: left hip, right hip,
left knee, and right knee. Each is driven by
a modified RC servo motor.
We constrain the robot to the
sagittal plane using a 1m-long
boom. The robot is attached
to the boom via a freely-rotat-
ing joint, and the boom to the
central column by a universal
joint. This boom structure
has negligible influence on
the dynamics of the robot in
the sagittal plane, allowing it
to freely trip or fall. The me-
chanical design of our robot
incorporates small curved feet
and a forward-located mass
center, both of which facili-
tate its fast-walking speed. It
also exploits natural dynam-
ics, such as inertia of the
limbs, friction of the motors,
and gravity.
The design of the neuromuscular con-
troller
The neuronal controller follows a hierarchi-
cal structure (see Figure 1). The bottom
level is the reflex circuit local to the joints,
including motor-neurons and angle sen-
sor neurons involved in the joint reflexes.
The top level is a distributed neural net-
work consisting of hip stretch receptors
and ground contact sensor neurons, that
modulate the local reflexes of the bottom
level. The effects of these sensor signals in
generating a walking gait are illustrated in
Figure 2. Neurons are modeled as non-spik-
ing neurons simulated on a Linux PC and
communicated to the robot via a DA/AD
(digital-analog/analog-digital) board.
2

We use a linear viscous elastic muscle
model that is composed of a spring in par-
allel with a viscous damper, and is directly
controlled by the motor-neuron output.
3

Each joint has an antagonistic muscle pair
of flexor and extensor, which are activated
by the extensor and flexor motor-neuron,
respectively (see Figure 1).
Figure 2. A series of frames of one walking step. At the time of frame 3,
the stretch receptor (Anterior Extreme Angle signal, AEA) of the swing
leg is activated, which triggers the extensor of the knee joint in this leg.
At the time of frame 7, the swing leg begins to touch the ground. This
ground contact signal triggers the hip extensor and knee flexor of the
stance leg, as well as the hip flexor and knee extensor of the swing leg.
Thus, the swing and stance legs swap their roles thereafter.
Figure 1. The circuit of the neuromuscular controller. Only the
muscle pair of one joint is illustrated.
Tao, continued p. 4 The Neuromorphic Engineer

Volume 3, Issue 1, May 006
A neurobiological
perspective on building
intelligent devices
The Neuromorphic
engineer
is published by the
Editor
Sunny Bains
Imperial College London
sunny@sunnybains.com
Editorial Assistant
Stuart Barr
newsletters@sunnybains.com
Layout Artist
Freddy B-Apeagyei
freddy@logikmedia.net
Editorial Board
David Balya
Avis Cohen
Ralph Etienne-Cummings
Timothy Horiuchi
Auke Ijspeert
Giacomo Indiveri
Shih-Chii Liu
Jonathan Tapson
Andr
é
van Schaik
This material is based upon work
supported by the National Science
Foundation under Grant No. IBN-
0129928. Any opinions, findings,
and conclusions or recommendations
expressed in this material are those of
the author(s) and do not necessarily
reflect the views of the National
Science Foundation.
The Institute of
Neuromorphic Engineering
Institute for Systems Research
AV Williams Bldg.
University of Maryland
College Park, MD 20742
http://www.ine-web.org
What is intelligence and, for that matter,
what problems might arise in building an
intelligent machine? Are human brains,
with their greatly expanded neocortex, the
only currently-existing intelligent devices?
Apart from some unconvincing computer
programs, the only known intelligent de-
vices seem to be animals: particularly birds,
and especially mammals. There is, however,
at least one other clear example of natural
intelligence: all living organisms, notori-
ously, appear to be intelligently designed,
even though this appearance is achieved
by selective amplification of molecular ac-
cidents. This form of natural intelligence
(i.e. the Darwinian algorithm comprised
of iterative replication/mutation/transcrip-
tion/translation/selection steps) is the only
other successful exemplar of intelligence
we have identified to date. It is also a good
source of clues to help us navigate the neo-
cortical labyrinth.
A good place to start our inquiry is to
ask what is going on inside the skull. (See
Figure 1) There are two basic processes:
the first is a rapid (millisecond) integration
step, in which synaptically weighted voltages
are collected over the surface of a neuron,
combined (possibly in a nonlinear way), and
sent via more synapses to other neurons.
There is also a slower learning process that
uses the rapid signals to modify the weights
such that performance improves. Learning is
done by adjusting the strength of individual
synapses according to the voltage across the
synapse (Hebbs Rule). The power of the
learned world model will reflect the extent
to which the synapses can individually be set
(much as the power of a digital computer re-
flects the number of transistors and memory
locations that can be individuallyand suf-
ficiently rapidlycontrolled.
Intelligence boils down to numbers:
the combinatorial potential vastness of the
world should be matched by a correspond-
ing potential combinatorial vastness of the
brain that models it, together with precise
rules (such as Hebbs) for selecting useful
combinations. Integration requires volt-
age spread, but accurate learning requires
chemical localization: the incompatibility of
these requirements limits intelligence.
The neocortex: looking inside the box
An enormous amount has been learned
about the neocortex. First, it seems to have
a similar microstructure in different animals
and different parts of the same animal, from
B
W
learning
genes
neocortex
W
Figure 1. The left hand picture illustrates the interaction between
an animals brain (B) and its world (W). The brains input-output
relation reflects its synaptic weights, which depend on the history of
ancestors (gray zone, genes) and, especially in complex animals,
on the history of the animal itself (learning). The right hand
picture shows the two main components of a mammals brain; the
subcortical structures (which learn pairwise correlations), and the
neocortex (which learns higher order correlations). The neocortex
relies particularly heavily on learning, and provides corrections to
subcortical computations.
Adams, continued p. 9
subcortical
structures
monotremes to Mozart.
The neocortex characteris-
tically has six layers. Neo-
cortical input arrives, from
a central and mysterious
lump of neurons called
the thalamus (layer 0),
in layer 4. The set of input
firings, filtered through
the 0/4 synapses, initial-
izes a representation that
then rapidly evolves as the
environment changes and
as inhibition and recurrent
excitation kick in. This
recurrent process may be
thought of as providing
a statistically optimal es-
timate of what the initial
pattern would have been
if there were no noise in
the neural circuitry.
1
Thus,
the core computation is,
as