creating robust manipulation interactions with imperfect actuators and sensors
DESCRIPTION
Creating Robust Manipulation Interactions with Imperfect Actuators and Sensors. Passive and active compliance with SEAs Highly integrated set of behaviors through behavior-based architecture Perceptual saliency amplification through efference-copy models - PowerPoint PPT PresentationTRANSCRIPT
Creating Robust Manipulation Interactions with Imperfect Actuators and Sensors
• Passive and active compliance with SEAs• Highly integrated set of behaviors through
behavior-based architecture• Perceptual saliency amplification through
efference-copy models• Exploit rich force interactions which naturally
occur during exploration as a learning opportunity
Passive and Active Compliance
Series Elastic Actuator Force based grasping
Exploiting Force Interactions for Learning
1. Exploration behavior1. Decrease shoulder
stiffness on contact2. Localizes exploration
around object3. Creates rich interaction
forces
2. Learning1. Force based
representation of object affordances
2. Model of natural interaction forces
3. Scaffold to richer manipulation abilities
Force Based Efference-Copy Model
Y
Z
X
m 1r
m 0r
m 0l
m 1l
0r
1r
2r
3r
4r
5r
0l
1l
2l
3l
4l
5l
1h
0h
6r
7r
8r
9r
6l
7l
8l
9l
Force Based Efference-Copy Model
• Predictive forward model of the joint torques
• Amplifies salient interaction forces during manipulation
• Torque predictions made using simple kinematic and mass model
Z n
Z 2
Z 1
EFF
+- EC
Exo
Ego
ExoGrav
MotCommanded torqueSensed torque
Predicted torque
Detection of Self-Induced Hand Forces
Interaction forces at hands are approximately equal and opposite
Interaction forces present
Interaction forces not present
Detection of Interaction Forces
Efference copy model generates torque prediction.
Torque prediction errors drivevisual attention system.
Ballistic reaching: prediction error
External forces: prediction error
Systems Development: Behavior Based Architecture
Arb
Wire
Overwrite
Max
Static
Arbitrator
Queue
Stack
Scheduler
Module
Thread
Monostable
•Architectural primitives allow tightly integrated system•100hz scheduler •Dynamic arbitration•12 node Linux cluster•~50 threads currently
Homeostat
VisualExploration
HandServoLeft
HandServoRight
HandLookRight
HandLookLeft
FaceTrack
BlobTrack
CartesianTrack
Fixation
VisualServo
Kinematics
ARB
PoseController TrackingController
ARB
BallTrack
ARB
I
ARB
Zero
I I I
FixationReach
CartesianTrack
ARB
ForceController VSpringController
ZeroG
ARB
VisualServoFingers
VisualServoProximity
I X
X
RelaxPose
ShowObject
StiffnessModulation
s
s
sx x
s
ExamplesArm Behaviors Head Behaviors
VisualExploration
HandServoLeft
HandServoRight
HandLookRight
HandLookLeft
FaceTrack
BlobTrack
CartesianTrack
Fixation
VisualServo
Kinematics
ARB
PoseController TrackingController
ARB
BallTrack
ARB
I
ARB
Zero
I I I
FixationReach
CartesianTrack
ARB
ForceController VSpringController
ZeroG
ARB
VisualServoFingers
VisualServoProximity
I X
X
RelaxPose
ShowObject
StiffnessModulation
s
s
sx x
s
ExamplesArm Behaviors Head Behaviors