Post on 25-Mar-2020
1
Jun UedaAssistant Professor, G.W.W. School of Mechanical Engineering
Adjunct Faculty, School of Applied Physiology
Georgia Institute of Technology
Modularity and Variability in Biologically inspired Actuation
Research topics[1] Biological (real) muscle research
� Exoskeleton for muscle diagnosis
� Human stiffness measurement
[2] Artificial muscle research� “Muscle-like” modular actuators
� Design of compliant mechanisms
2
[3] How can we merge
these two areas?
Generation of
natural movements
Understanding of
muscle coordination
Human sensorimotor
enhancement
Using Operator Arm Stiffness for ImprovedPerformance of Haptic Human-Robot Interfaces
� Problem: When instability is encountered, a human
operator often attempts to control the oscillation by
stiffening their arm, leading to a stiffer system with more
instability.
� Idea: Classify arm stiffness from EMG signals and choose
appropriate impedance parameters of a robot
3
Wearable EMG device
Problem: Instability of haptic interfaces� Power-assisting, or impedance controlled, lift device
� Increase of human stiffness� Co-contraction: Null space of muscle forces � joint-
toques (no change in end-point force)
4
Co-contraction of antagonistic muscle pairs
� Antagonistic muscles contract together � Higher stiffness
� Can calculate angle between moment arms
� Angles closer to 180°more antagonistic
� Calculated using exiting musculoskeletal model
Wrist (Flexor Carpi Ulnaris, Extensor Carpi Ulnaris) �127 degElbow (Biceps Brachii, Triceps Brachii) �163 deg
EMG measurement and stiffness classification
6
Measure agonist-antagonist muscle pairsWrist (Flexor Carpi Ulnaris, Extensor Carpi Ulnaris)Elbow (Biceps Brachii, Triceps Brachii)
0 5 10 15 20 25 30
EMG (BB)
EMG (TB)
EMG (ECU)
EMG (FCU)
Nom Angle
Nom Force
% of Variance of Stiffness
Pre
dict
or V
aria
ble
Predictor Variable Effect on Stiffness
2
7
Rigid Surface Contact Comparison
Stiffness compensation onStiffness compensation off
298 300 302 304 306 308
0.3
0.35
0.4
0.45
0.5
0.55
0.6
Time (s)
Pos
ition
(ra
d)
Rigid Surface Contact with Stiffness Compensation
Highlight indicates compensation for high arm stiffness
262 264 266 268 270 272 274 276 2780.3
0.35
0.4
0.45
0.5
0.55
0.6
Time (s)
Pos
ition
(ra
d)
Rigid Surface Contact without Stiffness Compensation
8
� Healthy participants (n=20)
� Average time of a pick-and-place
task was reduced by 25%
� High variability between subjects (and trials)
� Histogram of all measured stiffness levels shows a Poisson-type distribution
1 m
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Tim
e to
Bes
t Pla
cem
ent (
s)
(Nor
mal
ized
by
Dis
tanc
e to
Tar
get)
Controller State
Time (Normalized)
0
2
4
6
8
10
12
Dis
tanc
e of
Bes
t Pla
cem
ent f
rom
Tar
get
(cm
)
Controller State
Distance
Off
On
Performance test: Pick & Place Task
9
0 2 4 6 8 100
10
20
30
40
50
60
70
80
Stiffness (kN/rad)
Fre
quen
cy
Stiffness Level Histrogram
Im Is
VsVm
ZsZm
Um UsZenv
Zop
Vop
Operator Robot
Work
Operatorimpedance
Se
nso
r D
ata
Probabilistic estimation of muscle co-contraction: in-progress
10
Probabilistic prediction of
muscle co-contraction
S1 SnS2
1β
ImpedanceGain scheduler
2
0 argmin∑
=
j j
j
PCSA
f
ff
⋅<≤⋅=
jj PCSAkf0
fAτ
)()(0 tβAAIff +−+=
subject to
Hypothesis: (parameter of muscle coactivation) is a random variable & whoseprobability density function (PDF) can be predicted by means of Markov models.
)(tβ
11
Robot-assisted diagnosis of neurological movement disorders
� Exoskeleton robot can induce a wider variety of muscle activities than performing conventional tasks.
� Need to understand human-robot physical interaction at the level of individual muscles.
Exoskeleton has more “hands.”Control is more accurate.
12
Joint torques (e.g., by motion capture) Muscle forces
Optimality principles in redundant muscle coordination
1
( ) min
. .0
rn
j
j j
i
fu
PCSA
s tf k PCSAj
τ=
= →
= ⋅ < < ⋅
∑f
A f
Hypothesis : Brain coordinates redundant muscles to minimize a cost function
R. D. Crowninshield et al, J. Biomechanics, 1981.
1f3f
2f
1τ
2τ4f
5f
6f
Crowninshield’s law
fAτ ⋅=
=
2
1
ττ
:Muscle moment arm
:Muscle force
Af
SIMM, Musculographics inc. AnyBody , AnyBody, Tech. Ueda et al., 2007
3
Robot-assisted muscle isolation� Neuromuscular function test, therapeutic training, power-
assisting, muscle fitness training, …
� Can be boiled down to a single question:
13
How can we determine an adequate exercise that induces a desired change in a target muscle force?
13
Target muscle anddesired forces
Robot torque ??
Biceps = 15 [N] (x 1.5)
Biceps = 0.5 [N] (x 0.5)
or
orBiceps = 1.2 [N] (x 1.2)
min→
∑
j
r
j
i
PCSA
f
Tip-force ??
�Inverse solution of muscle force prediction based o n the optimality principle
Constraints for the cost function ?14
Theorem :Desired muscle forces can be realized if the exoskeleton applies
Individual muscle force control - Complete Solution -
(3) Inactive muscle forces are still zero
(1) Muscle forces are positive
(2) Desired muscle forces are completely realized
Feasibility conditions :
Target muscles are perfectly controlled
Minimize changes of non-target muscles
Jun Ueda, Ming Ding, Vijaya Krishnamoorthy, Minoru Shinohara, Tsukasa Ogasawara, "Individual Muscle Control using an Exoskeleton Robot for Muscle Function Testing,” IEEE Transactions on Neural and Rehabilitation Systems Engineering, 2010.
15
Force-matching motor tasks
Desired tip-force
Measured tip-force
Muscle control experiments
16
-0.6
-0.4
-0.2
0
0.2
0.4
BRA BRD
ECU
-0.6
-0.4
-0.2
0 0
BRA BRD ECU
Desired Muscle force
Modified Muscle force (EMG)
B:
BRA x 0.5BRD x 0.5ECU x 1.0A:
BRA x 1.0BRD x 1.0ECU x 1.3 C:
BRA x 0.5BRD x 0.5ECU x 1.3
Ratio of change
-0.4
-0.2
0.2
0.4
0 0
BRA BRD
ECU
Assist
Resist
0
(1) Elbow joint 90 deg(2) 2kg iron weight(3) 6 healthy subjects(4) Surface EMG measurement
Pneumatically-Powered Robotic Exoskeleton to Exercise Specific Lower Extremity Muscle Groups in Humans
• Exercise Anti-Gravity Muscles (Legs and Lower back) to shorten rehabilitation time when returning to a higher gravity environment
• Minimize Bone Losses since some loss may be permanent
Desired force profiles in atarget muscle
Research topics[1] Biological (real) muscle research
� Exoskeleton for muscle diagnosis
� Human stiffness measurement
[2] Artificial muscle research� “Muscle-like” modular actuators
� Design of compliant mechanisms
18
[3] How can we merge
these two areas?
Generation of
natural movements
Understanding of
muscle coordination
Human sensorimotor
enhancement
4
“Muscle” Actuator Materials and Biological Muscles
Stress
PZT SMA
Reliability,Stability
Efficiency Speed
Strain
PolypyrroleConductingPolymer
ElectrostrictivePolymer
(Elastomer)
0.1%, ~ 10ms
[2] Marieb, E. Human Anatomy and Physiology Benjamin Cummings, 2001
Compliant TissueCellular Structure (non-uniform)Quantized (on-off) controlVariability (noise) 19
“Nested Rhombus” Exponential Strain Amplification
PZT stack actuator
0.1%
1.6%
23.9%
1.6%
0.1% x 15 x 15=22.5%!
Our goal: 20% (Amplification gain)^Layerby “Power-law”
21% effective strain, 1.7N, 15g
20Ueda, Secord, Asada, IEEE/ASME Transactions on Mechatronics, Vol. 15, No. 5, pp. 770-782, 2010.
MRI-Compatible Self-sensing Piezoelectric Tweezers
Possible tele-surgery device• Nonmagnetic (MRI compatible)• 1N, >30Hz • Self-sensing of force/displacement
21
Self-sensing Piezo Tweezers
PZT
Differential amp Charge ampTrue valueSensor output
Actuator
Sensor
Yuichi Kurita, Fuyuki Sugihara, Jun Ueda, Tsukasa Ogasawara, "Piezoelectric Tweezers with Force- and Displacement-Sensing Capability for MRI, " IEEE/ASME Transactions on Mechatronics, in Press,
Amplifiedtip-force
Original tip-force
22
Bilateralcommunication
Nonmagnetic sensing system (Georgia Tech)
MRI compatible haptic device
MRI compatible fluid actuator (Vanderbilt)
23
3-spring static lumped parameter model
aloadk
Jk
pztf
pztk
1x∆BOk
BIkPZTk
loadk
PZT stackactuator
pztf
pztf
1x∆
1f1f
24
pztf
1f
pztx∆
1x∆ 2-port network model
∆∆
−−
=
−−
1132
21
1 / x
x
XXX
XX
f
f pztpzt
5
Nesting of two-port network models
25
� Outermost (n – k) layers dominate stiffness � Nesting Theorem
1I 2I
2V1V
1I 2I
2V1V
2I
2V
2I
2V
Load
Load
Unit 1
Unit 2
Unit n
PZT
2nd
laye
r am
plifi
catio
n m
echa
nism
1stla
yer
mec
han
ism
Act
uato
r ar
ray
with
nun
its
Schulz and Ueda, Invited Session- Bio-inspired Systems, Paper WeAT5.4, 11:00-11:20
Hill-type Muscle Model
Muscle-like compliance (Hill model??)
PZTf
PZTk
PZT stack actuator
Amplification mechanism
Lumped Model for Piezoelectric Cellular Actuators
26
Human Skeletal Muscle possesses:
� Quantization due to finite enervation rates
� Resonant modes due to flexibility and mass of muscle tissue.
However, despite these effects, humans are able to produce smooth motion with simple on-off commands to discrete groups of muscle fibers.
The biological mechanisms by which the switching times are chosen are not well known.
Physiologically-inspired impulse excitation method for fast and compliant actuators
Quantization in actuation (# of motor units) butHigh precision in time (very fast contraction time)
27
Vibration suppression by redundant ON-OFF actuation
0 5 10 15-1
0
1
2
3
4
Time [s]
Dis
plac
emen
t [m
m]
(1) Serially-connected PZT actuators (redundant discrete actuation)
(2) Fast response of PZT (high resolution in time)� Linear actuation (amplifiers) may not be necessary
Motivation
28ON-OFF
ON-OFF Power Switching Network
Step Minimum switching method (proposed)
Open loop-control (point-to-point movement)
Joshua Schultz and Jun Ueda, "Experimental Verification of Discrete Switching Vibration Suppression," IEEE/ASME Transactions on Mechatronics, in Press.
37% energy efficient
30
Biologically-inspired Robotic Vision
θ
tQua
ntiz
edC
omm
ands
“Saccadic” camera movement for quick panoramic composite photo
Object tracking by smooth-pursuit
Flexible structure (actuator)Quantized control
High-speed switching
6
31
Closed-loop control (smooth-pursuit)
0 0.2 0.4 0.6 0.8 1 1.20
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Time [sec]
Pos
ition
[m] Reference
PWM QuantizationIntersample quantized
0 0.2 0.4 0.6 0.8 1 1.2
Qua
ntiz
atio
n le
vel
Time [sec]
0
2
4
-2
-4
-6
-8
OFF
ON
Intersample quantized
Joshua Schultz and Jun Ueda, the 2010 Dynamic Systems and Control Conference (DSCC'10), Boston, MA, 2010.
0
3
1
2
T 2T
Nominal input (discrete-time control theory)PWM QuantizationIntersample Discretized
(1)
(2)
(3)
OFF
ON
Solving optimization problem
−−
1
1
3&
1
2
3&
1
2
3&
1
2
1&
L
L1 L2 Lk-1 Lk
L1 L2 L3 L4
−−
1
1
F1&
000
122
8&6&1&
000
211
01&E&1&
0
5
1&
L1 L2 L3 L4
L1 L2 Lk-1 Lk
Connection Fingerprint (programmable!)
(a)
(b)
“Fingerprint” method for modeling and characterizing complex actuator array topologies
David MacNair and Jun Ueda, "A Fingerprint Method for Variability and Robustness Analysis of Stochastically Controlled Cellular Actuator Arrays," The International Journal of Robotics Research, Volume 30, Issue 5, pp. 536 - 555, April 2011. 32
2 Cells : 2 Arrays
Automatic generation of actuator topologies using the fingerprint method
plot_fingerPrintGrid(fingerPrintBuild(# of cells)) 33
3 Cells : 4 Arrays
34
5 Cells : 23 Arrays
35
7 Cells : 199 Arrays
36
7
Non-uniform array response: twitch to tetanus
37
0
0.1
0.2
0.3
0.4
0.5
For
ce (
N)
0 5Time (s)
Cell 2, 3, 4Cell 8
Cell 10, 11
Cell 1, 6 Cell 5
Cell 7
Cell 9, 14Cell 12, 13
(a) single cell activations
0
1
For
ce (
N)
Time (s)
Cell #2 #2
#10#2
#10
#8
#11#3
(b) Multiple cell activations
11
2
3
4
5
6
7
8
9
10
11
12
13
14
Research topics[1] Biological (real) muscle research
� Exoskeleton for muscle diagnosis
� Human stiffness measurement
[2] Artificial muscle research� “Muscle-like” modular actuators
� Design of compliant mechanisms
38
[3] How can we merge
these two areas?
Generation of
natural movements
Understanding of
muscle coordination
Human sensorimotor
enhancement
39
ON-OFF ControlledSMA cellular actuators
Generation of “biological” movements
NSF CPS #0932208PI: Ueda, GT-ME
Musculoskeletal model
�����
Quantizer
ON-OFF switchingnetwork
Controller ����
)(tα )(tαTrajectory generator
Embedded sensor information
End-pointposition
Analog motor command
Quantized motor command
OFF
ONModulated impulse commands
Flexible muscle model
α
Physiologically-inspired quantized control
Henneman’s size principle?
Optimization criterion
min)()( →−∫+Rt
t d
f
f
dttt rr
• Non-uniform motor units• Achieving a fine resolution for a small
motor command and a more coarse resolution for a larger command
�
� �
�
(a) Uniform quantizer (b) Logarithmic (floating-point) quantizer
� �Neuronal variability?
41
Neuronal variability and Signal-dependent noise
Uno, Y., Kawato, M. & Suzuki, R. Biol. Cybern. 61, 89–101 (1989).Observation: human-arm trajectories, 4 trials
Signal-dependent noise (Harris and Wolpert, 1998): Standard deviation of force variability is proportional to mean motor command
Harris, and Wolpert, Signal-dependent noise determines motor planning, Nature, 1998
0 0.5 10
0.5
1
Motor command
Act
ivat
ion
leve
l
Criticisms:(1) Standard deviation of neurons’ firing rate is proportional to the mean rate to the power of 0.48.(2) What is the source of signal-dependent Gaussian noise?
Signal-dependent noise or quantization error ?
� Do we need a source of Gaussian noise in robots to mimic the
neuromotor variability?
� Henneman’s size principle �Floating-point quantization
� Our result: if an actuator array is floating-point quantized, its variability statistical is statistically equivalent to that of
proportional “signal-dependent” noise.
42
1 2 4 8
Activation level
Activation level
1 1 1 1 2 2 4 41 1 1 1 2 2 4 4
(a) floating-point segmentation with 1-bit mantissa (= binary segmentation)
(b) floating-point segmentation with 3-bit mantissa
8
Proof
43
Qx 'x
- +
FLν
compressorexpande
r
- +ν
Uniform quantizery 'y
FLQ Floating-point quantizer
Signal-dependent noise(Gaussian proportional to mean command)
Floating-point quantization
Equivalent if
0 0.5 10
0.5
1
Motor command
Act
iva
tion
leve
l
0 0.5 10
0.5
1
Motor command
Act
ivat
ion
leve
l
B. Widrow, Quantization noise, Cambridge University Press, 2008
44
0 0.2-0.2
0.5
0.4
0.3
Generation of biological movements: in-progress
Cellular actuator structure � Quantized noise in actuation
x [m]Uno, Y., Kawato, M. & Suzuki, R. Biol. Cybern. 61, 89–101 (1989).Observation: human-arm trajectories, 4 trials
min)()( →−∫+Rt
t d
f
f
dttt rr
Muscle-level comparison
45
Pec MajorBrachialisLatTriceps BrevBicepsTriceps Long
0 0.1 0.2 0.3 0.4 0.5 0.60
0.1
0.2
0.3
0.4
0.5
0.6
Time [s]0 0.1 0.2 0.3 0.4 0.5 0.6
Time [s]
0
0.1
0.2
0.3
0.4
0.5
0.6
Nor
mal
ized
mus
cle
activ
atio
n le
vel
0 0.1 0.2 0.3 0.4 0.5 0.6Time [s]
0
0.1
0.2
0.3
0.4
0.5
0.6
Harris’s proportional SDN Floating-point quantization Uniform quantization
� Proportional SDN, floating-point quantization, uniform quantization
-0.2 -0.1 0 0.1 0.2 0.30.2
0.25
0.3
0.35
0.4
0.45
0.5
X[m]
Y[m
] Proportional SDNFloating point quantized
Uniformly quantized
T2
T6
Sensorimotor Enhancer� Tactile receptors: nonlinear threshold systems � Stochastic resonance (application of subsensory
white noise) improves the sense of touch� Improve motor functions
46
Dr. Shinohara, Applied Physiology
Kurita, Shinohara, Ueda, ICRA 2011
Noise-enhanced human sensorimotor function
Khaodhiar, et al., 1991 H. R. Dinse et al., 2005
Collins, IEEE EMBC Mag. 2003
Challenges• Dexterity and force control of
fingers• Wearable (compact & palmar
side should be free)• Optimal design• Orthotic devices for persons with
peripheral neuropathy• Surgery training, Manual
assembly , texture design, etc.48
Experimental results
� Sensory test (Texture discrimination )
� Motor test (grasping test)
Higher sensitivity
Less effort(more efficient)
9
Transmissibility characteristics of a fingertip
49
Gain
Phase
• Vibration is attenuated by viscoelasticity of skin and subcutaneous tissue.
• Measured by Laser Doppler Vibrometry.
Challenges and opportunities
� Artificial systems and biological systems
� Quantized and compliant structure
� Modulated impulse commands
� Source of variability (noise)
�Wearable robotics
� Understanding of human-robot physical interaction at the level of muscles
� Dynamic and probabilistic modeling
� Sensorimotor systems and dexterity
� Evaluation of “human-like” movements?50
Acknowledgement� National Science Foundation (CPS-0932208, CNS-1059362, IIS-1142438)
� NSF Center for Compact and Efficient Fluid Power
� General Motors
� Korea Institute for Advancement of Technology
� Japan Science and Technology Agency
� RIM@GT, GT/Emory Health Systems Institute
51Thank you !!51
Dr. Yuich KuritaJoshua Schultz, MEDavid MacNair, RoboticsBilly Gallagher, RoboticsGreg Henderson, ME Melih Turkseven, ME Timothy McPherson, MEEllenor Brown, AP