Date post: | 20-Dec-2015 |
Category: |
Documents |
View: | 213 times |
Download: | 1 times |
Laboratory for Perceptual Robotics
Department of Computer ScienceUniversity of Massachusetts
Amherst
Natural Task Decomposition with Intrinsic Potential
FieldsSteve Hart Emily Horrell Shichao Ou Shiraj Sen John Sweeney Rod Grupen
Third Annual New England Manipulation Symposium
Rensselaer Polytechnic InstituteJune 1, 2007
Overview
• Control Decomposition– Intrinsic Potential Fields – The Control Basis Architecture– New Programming API
• What’s Next?– Multi-objective behavior– Generalization and Transfer
Potential Fields
- move-to- visual track- wrench-closure (grasp)
Extrinsic Navigation Functions
• Postural Bias- (range of motion)
• Kinematic Conditioning
Intrinsic Potential Fields
Kinematic Conditioning
• There are various traditional conditioning metrics for kinematic transformations.
Measure of Manipulability Measure of Localizability
Volume of “Conditioning Ellipsoid”
The Control Basis
• Combinatoric framework for sensorimotor control
: control function
: sensor
: effectorresourcemodel
(, ) : controller
: null space combinations
: sequential programs (schema)
∆2 1
• Control-based state/action space reflects the convergence status of a family of controllers
Dexter• The UMass BiManual
Robot
• Two 7-DOF Barrett Technology Whole-Arm Manipulators
• 2-DOF pan/tilt stereo head
• Two 3-Finger Hands with 6-axis load cell fingertip sensors
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
Control Basis API
Visual Inspection
QuickTime™ and aMPEG-4 Video decompressor
are needed to see this picture.
Conditioned Grasping
Classification Through Visual Conditioning
An Anticipated Developmental Programming Trajectory
Stage 2: touch what you see
Stage 3: learning “out of range”
Stage 1: learning about objects
• learning about objects:• stacking• seriation• humans
QuickTime™ and aCinepak decompressor
are needed to see this picture.
QuickTime™ and aYUV420 codec decompressor
are needed to see this picture.
QuickTime™ and aYUV420 codec decompressor
are needed to see this picture.
s0 s1 s2 s3a0 a1 a2
Schema Abstraction
procedural (how)
declarative (what)ai = i(i,i)
procedural (how)
declarative (what)s0 s1 s2 s3
a0 a1 a2
S0 S1 S2 S30 1 2
procedural
declarative
(0,0) (1,1) (2,2)
ai = i(i,i)
Schema Abstraction
s0 s1 s2 s3a0 a1 a2
procedural (how)
declarative (what)ai = i(i,i)
(1,1) (1,1)’
S0 S1 S2 S30 1 2
(0,0) (2,2)
declarative
defined by contextprocedural:
Schema Abstraction
Pick-and-Place Context Variations
QuickTime™ and aH.264 decompressor
are needed to see this picture.
Object Position= Left
LeftArmReach w/
RightArm Reach w/
Yes No
Object Scale= Large
No
Both ArmsReach w/
Yes
Transfer Learning - bringing prior experience to bear
Tele-operator sorting instruction
sorting replay w/ prior knowledge
(1) parse events to find
a matching schema.
(2) associate goalswith schema
(3) Replicatedemonstrationwith contingencies
QuickTime™ and aH.264 decompressor
are needed to see this picture.
QuickTime™ and aCinepak decompressor
are needed to see this picture.
Summary
QuickTime™ and aCinepak decompressor
are needed to see this picture.
• Control Decomposition
• Control Basis API
• Schema Learning
• Generalization/Transfer