Autonomous Robots **, *** ***, ****
c **** Kluwer Academic Publishers. Manufactured in The Netherlands.
Biologically-Inspired Collective Control for an Autonomous Robotic Arm
TYSON H. HARTY, GENE G. KORIENEK, CHARLES LEDDON AND ABIGAIL B. BAUTISTA
3 Sigma Robotics, 800 NW Starker Ave., Corvallis, OR 97330, USA
abpvj1@r.postjobfree.com
Abstract. Biological collective control architectures and simple control principles used in nervous systems pro-
vide novel alternative approaches for the design of fault-tolerant, adaptable real-world robotic systems that have
traditionally relied on centralized control. In this research, a robotic arm composed of multiple identical segments in
a collective computational architecture was tested for its ability to produce adaptive pointing and reaching behavior.
The movement rules for these robotic arm segments were derived from re ex arc principles in the human nervous
system. These arm segments received no central directions and used no direct informational exchange, but rather the
arm was sensor-driven at its leading segment in a way that maximized pointing accuracy of the arm. The remaining
non-leading segments in the arm were moved in a sequential order using only sensed locally-available movement
information about neighboring segments.
Pointing and reaching behavior was observed in experiments with and without obstacles to movement. Because
such behavior was not speci ed within each segment, the overall limb behavior emerged due to the interaction
and coordination of all segments, rather than due to any single segment, centrally controlled in uence, or explicit
inter-segmental method of communication.
Keywords: autonomous, collective, adaptive, emergence, nervous systems
1. Introduction To test this idea, we implemented a collective control
architecture in a multi-segmented robotic arm (Harty,
In biological and mechanical systems, manipulation 2000). Each segment in the arm contained its own in-
within a changing environment requires a form of con- dependent sensory, processing, and actuation compo-
trol that exhibits a large degree of adaptability and au- nents. The strategy of control for each segment was
tonomy. Biological limbs are very successful in this loosely derived from the re ex arc, the simplest way
environment while robotic and prosthetic devices have that vertebrate nervous systems control limbs during
yet to demonstrate a similar degree of exibility and precise yet adaptable movements (e.g., fundamentally,
adaptability. pointing and reaching) (Sherrington, 1947). We pur-
In order to design an autonomous adaptive robotic posely avoided using more complex control schemes
arm, we investigated one underlying reason for this bi- of the central nervous system and focused speci -
ological adaptability simple control principles in col- cally on the control principles observed in the pe-
lective computational architectures. We asked the ques- ripheral nervous system chaining, facilitation, and
tion: How can the individual components of a nervous inhibition as means by which multiple arm segment
system collectively generate group behaviors whose movements of the robotic arm could be coordinated
capabilities are greater than the sum of those individu- without direct centralized control. Facilitation and in-
als? One answer is observed in the emergent properties hibition occur when the movements of muscle groups
of these systems, due to the interaction of their many complement or oppose each other, respectively. For ex-
parts that together form group behaviors that cannot be ample, wrist exion facilitates elbow extension when
speci ed prior to their occurrence (Kelly, 1994). reaching for an object with the hand. Chaining occurs
300 Harty et al.
when the nervous system coordinates movement via plementation independence of both the control strat-
sequential relationships between muscle groups. Facil- egy and the collective computational architecture were
itative re ex chains, for example, are particular activa- tested.
tion sequences of muscle groups in various movements Each segment used plug-and-socket attachments to
such as reaching out one s arm to touch an object or dur- other segments, allowing rotation around its two con-
ing the righting re ex of mammals (e.g., dog, horse) nected axes. Thus, each segment had one axis that was
rising from lying to upright postures (Roberts, 1967). actively controlled and another that was passively
Thus, with the nervous system as a model, we de- turned via its attachment to an adjacent segment. Any
signed simple re ex control in a collective arrangement number of these single degree-of-freedom segments
of robotic segments with the objective of evaluating the could be connected to form a robotic arm of N degrees
resulting emergence of adaptive pointing and reaching of freedom.
toward a target in the presence of obstacles. The initial posture of the arm segments in every
experiment was a coiled posture with each segment
oriented at 160 to its proximally-adjacent neighbor.
This initial posture was chosen as a logical position
2. Methods
from which to initiate pointing and reaching behaviors
(Fig. 1). A vernier mechanism on each segment allowed
To focus on testing control algorithms rather than on
any two adjacent segments to be locked at relative po-
mechanical design, we used a puppet robotic arm, a
sitions to each other with a 1 angular resolution. The
linkage of identical segments that mechanically con-
arm was attached to one side of a 3 m cubic room and a
nected with one another. The segments did not contain
2D-Cartesian grid (100 cm by 100 cm) was placed as a
real actuators, sensors, controllers, or power sources,
target on the opposite wall. We mounted a laser pointer
yet were representative of a real robotic arm that could
on the end of the distal, or leading, segment of the arm to
contain these components. To facilitate moving to an
show the pointing location of the distal segment on this
actual electromechanical version, this physical proto-
target grid, allowing visual indication of pointing error.
type provided a useful 3D working model beyond a
To test for adaptability, an obstacle was placed be-
computer simulation.
The segments were identical molded plastic 90 tween the target and the arm to inhibit movement. The
obstacle was a 33 cm diameter circle mounted through
toroidal arcs 10 cm in length (Fig. 1). The segment
its center on a tripod and chosen based on its rela-
and overall arm shape were not designed to be direct
tive size to the segments. Fifty experiments were con-
anatomical analogs of human arm joints, but instead
ducted, based on a combination of: (1) segment num-
were designed to test principles of control in an ab-
ber (3 7), (2) presence or absence of obstacles, and
stracted mechanical form outside the morphological
(3) control rules (Table 1). The initial posture of the
constraints of the biological realm. In this way, the im-
arm segments and target location were constant in all
experiments.
For control rules, we operationalized re ex arc prin-
ciples into conventional production rules using the gen-
eral form:
IF: Distally-adjacent segment direction is [CW
or CCW] and rotation = [x ],
THEN: Rotate [transformed x ] degrees in a [trans-
formed] direction.
Facilitative chaining (FC) was used when a given
arm segment maintained the rotational magnitude and
direction of its adjacent segment and inhibitive chain-
ing (IC) when the given segment opposed that rotation.
In addition, we used an analog of the nociceptor re ex
(i.e. the re ex which causes one to pull one s arm away
Figure 1. A 7-segment robotic arm attached to mounting plate in
if a hot object is touched) as a withdrawal re ex (WR)
initial posture.
Biologically-Inspired Collective Control 301
Table 1. The control rules used in each segment of the collective robotic arm.
Facilitative Chaining (FC):
Distally-adjacent segment moves CW (or CCW),
IF:
THEN: Rotate ( 0.667) in the same direction.
Inhibitive Chaining (IC):
Distally-adjacent segment moves CW (or CCW),
IF:
THEN: Rotate ( 0.667) in the opposite direction.
Withdrawel Re ex #1 (WR1):
Distal Segment Rule:
IF: CW or CCW rotation while attempting to point as close as possible to the target is inhibited by an obstacle or another segment,
THEN: Reverse direction and attempt to point to the target.
Non-Distal Segment Rule:
IF: A reversal in direction by distally-adjacent segment is sensed,
THEN: Movement magnitude = [0.667 (Y X )], where [X = magnitude of the distally-adjacent segment s rotation before the
reversal], and Y = magnitude after the reversal]; Movement direction = the opposite of as that of the distally-adjacent
segment s direction before the reversal.
Withdrawel Re ex #2 (WR2):
Distal Segment Rule:
IF: CW or CCW rotation while attempting to point as close as possible to the target is inhibited by an obstacle or another segment,
THEN: Rotate Y in the opposite direction, where Y = 20 .
Non-Distal Segment Rule:
IF: A reversal in direction by distally-adjacent segment in sensed,
THEN: Movement magnitude = [0.667 (X + (0.20) X )], where [X = the magnitude of the distally-adjacent segment s rotation
before the reversal]. Movement direction = the opposite of that of the distally-adjacent segment s direction before the reversal.
control strategy in the chain of segments. This WR discrete segment move to the maximum reaching dis-
strategy was used by segments to initiate a rule switch- tance of the outstretched arm. All data was recorded
ing behavior when triggered by a direction-switching in a computer spreadsheet that computed performance
movement of an adjacent arm segment (Sherrington, measures and all segment movements for the human
1947). experimenter.
Different rules were used for the distal and non-distal All experiments began with an arm in the xed initial
segments, since the distal segment used pointing error posture. The human experimenter rotated the distal seg-
feedback from the target, while the non-distal segments ment in either a clockwise (CW) or counterclockwise
used locally-available sensed rotational movement (CCW) direction (relative to its neighboring segment)
magnitude and direction from their distally-adjacent to minimize the APE. The nal magnitude and direc-
neighboring segments. All non-distal segments applied tion from initial posture to nal posture of the distal
the same control rules in a given experiment. The rules segment were recorded and used as sensory input by
also used a two-thirds (0.667) attenuation of movement the next segment. Next, each remaining non-distal seg-
magnitude between segments that effectively balanced ment was moved sequentially in a distal-to-proximal
a segment s position in the arm. This relationship was fashion according to the control rules that transformed
chosen because it has been observed for relative angular the movement of its distally-adjacent neighboring seg-
velocities between joints in simple human movements ment. After each move, each segment was locked in
(De Sperati and Stucchi, 1995). place using its vernier mechanism. Each experiment
For each experiment, two dependent measures were ended when the distal segment was no longer able to
recorded after every segment s movement: Absolute make any additional movements beyond its current po-
Pointing Error (APE) and Reaching Error Ratio (RER). sition to minimize its APE. For all experiments in this
APE was the distance from the pointing location (x, y ) study, segment movements were conducted in an or-
of the distal segment to the target origin (0, 0). RER dered, distal-to-proximal sequence.
was the ratio of the distance of the tip of the laser In experiments with obstacles, the obstacle center
pointer on the distal segment to the target after each was aligned with the axis between the arm mounting
302 Harty et al.
plate and the target origin, with the obstacle face
perpendicular to this axis. The obstacle was placed
halfway between the arm s maximal reach and its initial
posture. This obstacle placement effectively blocked a
direct pointing trajectory, designed so that the arm had
to nd an alternate pointing solution to the target.
3. Results and Discussion
Thirteen out of the 50 experiments ended with a nal
pointing location on the target grid. Seven of these 13
experiments had a nal APE of 25-cm or less. Two out
of those 13 experiments ended exactly on target (i.e. Figure 3. Example plot of Absolute Pointing Errors after each seg-
APE = 0). One of these was a 4-segment arm without ment move for a 3-segment robotic arm, demonstrating the successful
coordinated point-to-target behavior.
an obstacle present that used a FC strategy. The second
was a 3-segment arm with an obstacle present that used
FC and WR strategy #1.
All experiments that ended with pointing solutions
on the target grid used searching patterns in their
segment-by-segment movements (Fig. 2). These search
patterns did not show a smoothly decreasing pro-
gression, but rather a choppy movement due to the
90 relationships between segments. However, if the
pointing locations following each complete arm move-
ment, rather than segment movement, were isolated, a
smoothly decreasing trend emerged (i.e., dotted line
in Fig. 2). Likewise, search patterns were evident in the
decreasing progression of APE from the initial posture
to the nal pointing solution of the arm (Fig. 3).
Figure 4. Plot of nal Reaching Error Ratio, demonstrating
In terms of reaching, 8 out of 36 experiments ended
reaching-to-target behavior for a robotic arm with 4 segments, using
with nal RERs of 50% or greater, that is, they reached
a Facilitative Chaining strategy.
at least half the distance they could maximally reach
to the target. One of the experiments that had a nal
APE of zero also had the largest nal RER of 0.86. This
experiment used an arm with 4 segments and no obsta-
cle present and a FC strategy. Across all experiments,
it was observed that reaching behavior could be pro-
duced over successive distal-to-proximal movements
(Fig. 4).
Across all 50 experiments, the mean nal APE in
experiments with no obstacles present (n = 10) was
33-cm, compared to 46-cm for experiments with obsta-
cles (n = 23). Similarly, mean nal RER in experiments
with no obstacles present was 0.51, compared to 0.12
for experiments with obstacles. Thus, while control
Figure 2. Plot of pointing positions on target grid for a 3-segment
rules in experiments with obstacles were less produc-
robotic arm, showing searching behavior before focusing the distal
tive as a whole than in experiments without obstacles,
segment on the target grid center.
Biologically-Inspired Collective Control 303
the fact that these control rules were able to generate the overall movement goal of the arm as a whole,
new movement solutions and adapt in some degree to it was informationally isolated from all other seg-
the obstacle shows promise for the use of simple mo- ments. Therefore, each segment acted as a single pur-
tor control rules in a collective architecture to produce pose machine encapsulating a speci c adjustment to
adaptive obstacle-avoidance behavior. environmental change. Yet from this direct relation-
In terms of control rules, the FC strategy promoted ship between individual segments and their environ-
distal segment orientation via the successive move- ments, pointing and reaching in the arm as a whole
ment of non-distal segments when no obstacles were occurred.
present. However, in the presence of obstacles, FC did (2) A form of emergent group behavior was ob-
not promote adaptation, since the movement of non- served. The robotic arm did not use centrally driven
distal segments facilitated further movement into commands (e.g., reach to the target there ), but rather
the obstacle once the distal segment had encountered was sensor-driven by the distal segment s pointing
it. Thus, while FC pointed successfully without ob- error feedback. However with its single degree of free-
stacles, it became clear that additional control rules dom of movement alone, the distal segment could not
were needed to promote adaptation to obstacles. As achieve an accurate pointing solution in most cases.
such, the IC and WR strategies did not drive the arm Only through cooperation with other segments move-
further into an encumbered situation, but tended to ments could the distal segment s own limited interac-
back the arm out of such situations using direction- tion with the environment be used to guide the arm to
reversing behavior. Thus, when an obstacle was en- a successful pointing solution. The arm s control so-
countered, both these strategies gave the arm more non- lution was not embedded in a single segment, but was
obstacle-encumbered working space than did the FC completed by the interactions of every segment. In this
strategy. sense, there was no need for a central in uence for
In fact, the WR strategies (i.e., ones not using pure each segment s movement. Rather, the distal segment
FC or IC) used a FC rule as their fundamental behavior- was modulated, via its pointing sensor, by informa-
generating strategy, but employed a rule switching tion embedded in the environment (i.e., some form of
tactic when obstacles were encountered. In effect, this emitted target energy ).
rule switching used a certain pattern of perceived sen- Additionally, no speci c reaching feedback infor-
sory information to choose a different subset of the mation was made available to any of the segments,
control rule. This enabled the arm segments to adapt and no speci c parameters were incorporated into the
to obstacles by sensing barriers to movement and control rules for reaching as there were for point-
responding by reversing direction. Neighboring seg- ing. Yet, reaching behavior was observed for the
ments sensed this non-explicit form of communication arm.
as the WR pattern of reversal movement of their neigh- Thus, pointing and reaching were the sum of the
boring segment, and made appropriate rule switching collective movements of all segments in the robotic
decisions to control their own movement. arm. This group behavior clearly was not speci ed in
the control rules within each segment, yet the arm as a
4. Conclusions whole was able to produce these coordinated behaviors.
This lack of task speci cation from a global perspective
Though not all possible obstacle placements, target lo- suggests that pointing and reaching emerged from the
cations, and initial postures for this robotic model were individual contributions of each segment.
tested, the ones used were a subset from which we were However, can this robotic arm s behavior be emer-
able to draw some conclusions about the usefulness of gent in the same sense that biological behavior is
re ex arc principles of motor control and a collective emergent ? This robotic arm consisted of multiple
control architecture in a robotic system: interacting deterministic processes that were tested us-
(1) Simple rules used to control the local move- ing sequential control, rather than controlled as mul-
ments of individuals in a collective computational tiple non-linear processes moving asynchronously. In
architecture can produce coordinated group behav- a biological system, emergence occurs because of
ior. Each segment could sense only its own interac- the high degree of non-determinism in its pieces and
tion with the environment and control only its own their interactions (Kelly, 1994; Holland, 1998). Perhaps
actions. Since each had no shared representation of then, this robotic arm lies somewhere in the middle
304 Harty et al.
of the spectrum of behavior between speci ed and Acknowledgments
emergent .
In sum, this research provided evidence that adapt- This work was supported in part by a National
able pointing and reaching movements in a robotic arm Science Foundation Graduate Research Fellowship to
with a collective control architecture might be pos- T.H. Harty.
sible without explicit programming. Such a robotic
arm has applications in remote and hazardous envi- References
ronments and as a new breed of prosthetics. The space
program, in particular, could use such an arm in space- De Sperati, C. and Stucci, M. 1995. Visual tuning to kinematics
craft manipulators or for planetary surface exploration. of biological motion: The role of eye movements. Experimental
Brain Research, 105:254 260.
The style of exible, adaptable control exhibited by
Harty, T.H. 2000. The application of human motor control principles
such a robotic arm would contrast with the algorithm-
to a collective robotic arm. M.S. Thesis, Biological Control Lab.,
based, precise, and highly repeatable control that is Oregon State University.
the trademark of conventional robotic arms. Yet, these Holland, J.H. 1998. Emergence: From Chaos to Order, Addison-
two styles of control produce a tradeoff in operation Wesley: Reading, MA.
Kelly, K. 1994. Out of Control: The Rise of the Neo-Biological
and behavior, demonstrating that robots with novel
Civilization, Addison-Wesley: Reading, MA.
biologically-inspired control systems could ll appli-
Roberts, T.D.M. 1967. Neurophysiology of Postural Mechanisms,
cation niches in which conventional robots are limited, Plenum Press: New York.
especially those in which unpredictable, dynamic con- Sherrington, C.S. 1947. The Integrative Action of the Nervous System,
ditions are the mainstay. 2nd ed., Yale University Press: New Haven.