Domestic Rehabilitation and Learning of TaskSpecific Movements

1
Guiding questions
What strategies are used in insect locomotion and
what are their implications?
MURI
Low-Level Control
Insect locomotion studies (Berkeley Bio) New
measurement capabilities (Stanford)
What motor control adaptation strategies do
people use and how can they be applied to robots?
Fabrication
Learning and Compliance Strategies for
Unstructured Environments (Harvard Johns
Hopkins) Implications for biomimetic robots
(Harvard, Johns Hopkins, Stanford)
2
Biological Motor Control
aero-, hydro-, terra-dynamic
Higher
Sensors
Environment
Centers
Preflexes
Open-loop
Mechanical
Feedforward
Behavior
System
Controller
(CPG)
(Actuators, limbs)
Feedback
Closed-loop
Controller
Adaptive
Sensors
Controller
3
Human Arm Model

• Simplifies experiments
• Excellent adaptability
• Instructable subjects
• Simple apparatus
• Manipulation application
• Role of impedance less understood

4
Learning Impedance Strategies in Unstructured
Environments

• Robert D. Howe and Yoky Matsuoka

Division of Engineering and Applied
Sciences Harvard University
5
Contribution to Control
Mechanical System
Feedforward
Preflex
Intrinsic musculo-skeletal properties
Motor program acting through moment arms
Predictive
Rapid acting
Passive Dynamic Self-stabilization
6
Understanding Impedance Change over Time

• Impedance preflex can produce robust behavior
  (Full)
• Preflexes are tailored to specific tasks and
  environments
• Goal Understand relationship between impedance
  value and task/environment
• Approach Measure impedances and adaptation
  strategies in realistic settings

7
Experimental Technique

• Develop a system that identifies impedance during
  execution of various tasks.
• Virtual Environment
• Characterize impedance change over time
• Instantaneous identification technique
• Investigate impedance adaptation characteristics
• How do humans adapt to a required impedance for
  the task?
• What is the initial strategy for a novel task?
• What does the initial strategy depend on?

8
Assumptions for System Identification
F

• Work with the end-point impedance
• Represent hand as a linear, second-order system.

m

• Parameter identification is easy for time-
  invariant systems
• - Assume constant m,b, and k
• - Apply perturbation, repeat, and average.

b
k
9
Previous System Identification Technique for Time
Varying Systems

• Time varying system
• Requires multiple perturbations for each data
  point.
• PRBS (Bennett et al. 1992, Lacquaniti et al.
  1993)
• Repetition hides learning
• Single task

10
Creating a Task-Based Environment

• Use a PHANToM haptic interface (3 DOF) to apply
  task-based force feedback
• Permits software control of task parameters
• Use force and acceleration sensors near the hand.

Robot
Force sensor
Handle accelerometer
11
Controller Interactions
Motors and Encoders
Robot
Force sensor
Data Acquisition System (10kHz)
Handle accelerometer
Processor (servo loop 1kHz)
Virtual Environment Dynamic Simulation
Computer Monitor (30Hz)
Human Subject
12
Using Impulse-Based Instantaneous System
Identification

• Use task-based force feedback if task interaction
  is impulse like
• Use added impulse force perturbation otherwise
• Identify within 40 msec, prior to CNS
  involvement - prefelexes only
• Assume passive impedance is constant during 40
  msec identification window

13
Example Task Bouncing a ball to a target height
VIEW ON MONITOR
Before During After
Handle corresponds to the paddle on the monitor
Ball drops too quickly for visual reaction
bounce height set by hand impedance
14
Contact Task Model Three Stages in Bouncing a
Ball

• Stage 1 Ball falling (before contact)
• given
• Stage 2 During contact
• Stage 3 Ball rising (after contact)
• given

ball
15
Task-Based Impulsive Force
Using the virtual environment contact task force
16
Confirmation of the Technique

• Least Square Fit
• r 0.988 (mean)
• Accuracy
• mass /- 8.1
• spring /- 2.5

2
1.5
1
0.5
ma
2
kx
0
Force (N)
-0.5
bv
-1
-1.5
F
-2
-2.5
-3
0
10
20
30
40
Time (msec)
17
Contact Task Experimental Design 1

• Used wrist movement to bounce the ball up to a
  target height
• Impedance too high bounces too high
• Impedance too low bounces too low
• Visual feedback of success/failure each trial
• Six sets of 40 bounces
• Group1 low, high, , low, highGroup2 high,
  low, , high, low

18
Typical Stiffness Learning Curve (n1)
High Target
Low Target
Group2 (high, low)
K (N/m)
High Target
Low Target
Group1 (low, high)
trials
19
Typical Damping Learning Curve (n1)
High Target
Low Target
30
30
25
25
Group2 (high, low)
20
20
15
15
10
10
0
10
20
30
40
0
10
20
30
40
B (N.s/m)
Low Target
High Target
30
30
Group1 (low, high)
25
25
20
20
15
15
10
10
0
10
20
30
40
0
10
20
30
40
trials
20
First versus Second Exposure to the Task (n5)
Low Target
High Target
1000
1000
800
800
First Exposure
600
600
400
400
200
200
A - 0.08
0
0
0
10
20
30
40
0
10
20
30
40
K (N/m)
High Target
Low Target
1000
800
600
Second Exposure
400
200
A - 0.3
0
0
10
20
30
40
0
10
20
30
40
trials
21
Stiffness Adaptation Characteristics for Contact
Task Experiment

• Initial stiffness same (200 N/m) regardless of
  target impedance
• Final stiffness tuned to target (range 200-650
  N/m)
• Learning follows exponential curve
• Adaptation is faster for the second exposure (for
  high target impedance)
• - First exposure A - 0.08
• - Second exposure A - 0.3

22
Precision Limitations
NARROWING TARGET STIFFNESS
NARROWING TARGET DAMPING
23
Precision limitations
Success rate with narrowing window
Average Stiffness
Average Damping
50
25
0
1
2
3
4
5
6
NARROWING TARGET WINDOW
24
Position Task Experimental Design

• Used wrist movement to track a moving ball
  (const. velocity).
• Game over if ball dropped.
• 5 continuous minutes of recording with added
  perturbations.
• Group1 Large paddleGroup2 Small paddle

Handle corresponds to the paddle on the computer
screen
25
Stiffness Adaptation in Position Task (n5)
Group 2 Small paddle
Group 1 Large paddle
K(N/m)
K(N/m)
time (1/10 min)
time (1/10 min)
26
Stiffness Adaptation Characteristics for
Position Task Experiment

• Initial stiffness same (740 N/m) regardless of
  paddle size.
• Amplitude of initial stiffness different for
  position and contact tasks.
• Final stiffness for the easier task, stiffness
  dropped lower and faster.

Group1 Large paddle
Group2 Small paddle
27
Summary

• Developed new experimental technique
• - instantaneous impedance measurement
  permits examination of learning and adaptation
• - virtual environment allows easy examination of
  a wide range of tasks
• Initial strategy depends on the overall task
• Final strategy depends on the environmental
  parameters
• Damping cannot be independently controlled