Recognizing Operator Intention via control basis projection: Early Experiments

1
Recognizing Operator Intention via control basis
projection Early Experiments

• Gary Holness, Sichau Oh, Robert Platt,Dr.
  Michael Rosenstein, Andrew Fagg

2
With actuators we get the robot to move
3
Controller Generate Motion

• Control law maintains model of the robot -
  physics masses, moments, friction etc -
  geometry link lengths, joint types, kinematics
• - model for moving the mechanism

4
Teleoperator control


Torso Teleoperation

• Serves as control input
• Drives the arm

5
Recognizing operator intent

• Whose head did operator intend to crush
• Explain task by projecting onto control basis

?
6
Explaining tasks with canned control dynamics

• Sample the motion path
• trajectory velocity vector between point and its
  sucessor
• Controller computes commanded velocity
• Closest one is best controller

Trajectory

• Later What about Null spaces?

7
2 Methods
8
Approach

• Goals local minima in velocity space
• Instantiate controller parameterized with each
  goal
• Compute trajectory velocity vector
• Controller compute commanded velocity vector

9
First experiment

• 6 Goals 3 x move to home??park
• Assumption pause at goal
• Time steps 220
• cos ?
• L(?)
• goal(0) 1.569133 1.562885 0.000449 0.018949
  0.006588 -0.018636 0.010991 //park
• goal(1) 0.014451 0.003306 0.007037 -0.006686
  0.006904 -0.016267 0.011711 //home
• goal(2) 1.568870 1.563108 0.000823 0.018949
  0.006430 -0.018162 0.011094 //park
• goal(3) 0.014451 0.004680 0.006788 0.004098
  0.007693 -0.014214 0.011711 //home
• goal(4) 1.568695 1.562828 0.000171 0.018772
  0.007298 -0.018399 0.011197 //park
• goal(5) 0.014451 0.007650 0.006662 0.011170
  0.007693 -0.009792 0.011711 //home

10
(No Transcript)
11
(No Transcript)
12
Future considerations

• HMM approach for recognition
• Interactive goal recommendation project forward
  in time toward goal
• Null-spaces descriptive power