ieee css 2015 04-03
TRANSCRIPT
Control of Human Movement:
from Physiology to Engineering
Antonie J. (Ton) van den Bogert
Parker-Hannifin Endowed Chair in Human Motion and Control
Department of Mechanical Engineering
Cleveland State University
http://hmc.csuohio.edu
IEEE Control Systems Society, Cleveland, 4/3/2015
Cost of transport (COT)
distance x weight
used nergy eCOT
COT = 0.2 COT = 3.0
Non-human movement: Honda Asimo
Petman (Boston Dynamics)
cost of transport unknown -- designed for tethered operation
Big Dog (Boston Dynamics)
3 MPG
340 MPG
hydraulic actuators powered by
internal combustion enging
Indego Exoskeleton (Parker-Hannifin)
electric motors powered by rechargeable batteries
Humans and animals
Humans and animals use 10-100 times less energy than machines
perform much better than machines
Why?
What can engineers learn?
Physiology (sensing/actuation/control)
Structure of muscles
2 μm
Overlapping proteins in muscle fiber
ADP
actin filament
myosin filament &
myosin head (crossbridge)
ATP
ATP (adenosine tri-phosphate) is
energy source of muscle
contraction
10 nm stroke length
Control of muscle force
Luigi Galvani (1737-1798)
Hz
Hz
Hz
Hz
twitches
fused tetanus
Frequency-modulated pulse trains
Motor unit recruitment
http://nmrc.bu.edu/tutorials/motor_units/
Motor unit:
A set of muscle fibers that are controlled by the same motor neuron
When force increases gradually, smallest motor units are recruited first
Spring-like mechanical properties
fiber length
forc
e
Control system needs to know
that force is position-dependent
Velocity dependence of force
Muscle (like any motor) has an optimal speed of operation
typically about 0.3 m/s (depending on muscle architecture)
Muscles vs. electric motors
Muscles
50 ms response time
slower to turn off
20-25% efficiency
low speed
high torque
"direct drive"
Electric motors
instantaneous response
90% efficiency
high speed
low torque
requires gearbox for
human-like applications
Feedback control
Physiological sensors for motion control
skin (stretch and pressure)
inner ear (inertial sensors)
muscles (stretch and force)
Nerve conduction velocity is about 100 m/s
Reflex loop delay 50 ms
Animals vs. machines Muscles
slow (50 ms response time)
inefficient (20-25% efficiency)
inconsistent (fatigue, variability)
Sensory system slow (50 ms signal delay)
inconsistent
Sprint running: foot is on the ground for only 100 ms!
Why do humans and animals perform so well? mechanical design (anatomy)
control (brain and spinal cord)
Anatomy (mechanical design)
Muscles often cross multiple joints
motor
motor
arm bones and muscles
typical robotic design
0 0.5 1 1.5 20
0.5
1
1.5
2
2.5
LENGTH
FO
RC
E
passive
25% activated
50% activated
75% activated
100% activated
Muscles are like springs
• Nonlinear spring
• Activation moves you to a different force-length curve
isometric
isotonic
Horse limbs
muscle
tendon
This makes control easier and saves energy
Muybridge, 1878
A spring-based exoskeleton
23
www.cadencebiomedical.com
What about motors?
Hanz Richter, CSU Robotics Lab:
"Semiactive virtual control"
Motors can
transfer energy
store energy
Energy-efficient robots
courtesy of Sangbae Kim, MIT
MIT Cheetah robot
uses less energy than animal
at same size and speed
Control (brain and spinal cord)
Hierarchical system
Brain: 1011 neurons
Spinal cord: 109 neurons
~10,000 PCs
Proportional-derivative control
Seems to be used by humans for simple
movements (reaching)
Also known as
Equilibrium point control (human motor control)
Impedance control, compliance control (robotics)
position
force
actuator with
elastic properties or
proportional control
external load
Control of standing Proportional-derivative control works well for small
perturbations
ADRC works well also
Larger perturbations require stepping move away from the desired posture!
proportional control will never do that
𝐱 =𝜃𝑎𝑛𝑘 𝜃𝑎𝑛𝑘𝜃ℎ𝑖𝑝 𝜃ℎ𝑖𝑝
u =𝑇𝑎𝑛𝑘𝑇ℎ𝑖𝑝
= −𝐊2x4𝐱
Walking is even more complex
Nonlinear dynamics
High-dimensional state space and control space
Limit cycle
Proportional control is not always "smart" enough
1650 RR uxu)f(x,x
Proportional-derivative control
designed by linearization
Muscles receive feedback from joint angles and angular velocities
Simulation test Human response to tripping
Do we need different control laws?
Do we need additional sensors?
Identification of human control
Human-based control:
We "map" the control system of our
volunteers, so we can copy it to a
robotic system
gain-scheduled
PD control
neural networks?
Virtual muscles
Electric motors can behave and feel like real muscles
motor
motor
=
joint
rotations
joint
torques
Summary
Animals and humans can perform amazing movements
more efficient than most robots
Muscles and nerves inefficient, slow and sloppy
Mechanical design is important can be virtualized with electric motors
Control is important learn from human data
Thank You!