Download Genetic algorithms for dexterous manipulation

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Philopatry wikipedia , lookup

Epistasis wikipedia , lookup

Population genetics wikipedia , lookup

Gene expression programming wikipedia , lookup

Transcript
Genetic algorithms for dexterous
manipulation
Seminar in Machine Learning
June 2005
Outline
•
•
•
•
•
Dexterous manipulation
Robot hands
Genetic algorithms and Genetic programming
Biologically inspired robot grasping using GP
A biologically inspired fitness function for robot
grasping
• Conclusions
Dexterous manipulation
The capability of changing the position and orientation of the
manipulated object from a given reference configuration to a
different one, arbitrarily chosen within the hand workspace
• Dexterous manipulation can be used to increase
flexibility and provide richer manipulation patterns
Robot hands
• Reliability and compactness and the developments of
suitable sensors, in particular effective fingertip
sensors
• Grasp synthesis/Grasp planning
• Grasp stability - how many fingers are necessary in
order to stably grasp a given object and where should
these fingers be placed under various conditions
Robot hands
• Form and Force Closure
Form Closure: the ability of a grasp to prevent motions of the
object relying on only unilateral, frictionless contact
constraints
Force Closure: the situation where motions of the object are
constrained by suitably large contact forces of the grasp
(usually considering friction)
GA and GP
• Created by John Henry Holland in 1960’s
• AI method that mimics the natural biological
processes
• Transforms a set of individual mathematical objects
by using the Darwinian principle of survival of the
fittest
Crossover, genetic recombination
Mutation
Reproduction
GA/GP for dexterous manipulation
• Initial population of candidate grasps
• Each member of the population (grasp) is evaluated
for its fitness
• The GA evolves the population through a number of
generations of candidate grasps
• The evolution is guided by a fitness function provided
by the user
• In the end a final superior grasp population is
generated
Biologically inspired robot grasping
using GP
Robot grasping
• Use of a GA to solve the grasp synthesis problem for
multifingered robot hands
• Select a “best” grasp of an object given some
information about the object geometry and some
user-defined fitness functions
• Fitness function is used by the GA to evolve
populations of candidate grasps
• Provide the operator with a general grasp selection
planner which can select and preview candidate
grasps across a wide range of objects and tasks
Robot grasping
• Example with a softball to be grasped by the threefingered Barrett Hand
• The program begins with a description of the object to
be grasped and an open hand
• The program placed imaginary spheres onto the
fingertips of the hand
• Intersection between the spheres and the object are
later used to evaluate when the fingers are in contact
and also to adjust the tightness of a grasp
• The program positions the object roughly central to
the workspace of the hand
Robot grasping
• Barrett hand with object
Robot grasping
• Grasp population member and evolution to one-finger
contact
Robot grasping
• Evolution to two-finger and three-finger contact
Robot grasping
• Final grasp
Robot grasping
• Grasping procedure
A biologically Inspired Fitness
Function for Robot Grasping
Fitness Function for Grasping
• Solutions represented in a tree format
• The root node represents the entire hand
• Each of the node that branches out represents a
different finger
• The finger nodes have three additional nodes that
represent the position of each joint
Fitness Function for Grasping
• Representation trees
Fitness Function for Grasping
• Mutation
Fitness Function for Grasping
• Crossover
Fitness Function for Grasping
• Virtual spheres
Fitness Function for Grasping
• Triangle areas and triangle angles
Fitness Function for Grasping
• Multiple object planes
Fitness Function for Grasping
• Raccoon grasping
Fitness Function for Grasping
• Finger intersections
Fitness Function for Grasping
• Pseudo equation for the fitness function
Fitness function =
Area of intersection of virtual spheres (sum in square inches) +
Triangle area (in square inches) +
Triangle angles (sum in radians) +
Number of contact planes (Maximum=9) +
Angle between fingers and object (sum in radians)
A biologically Inspired Fitness
Function for Robot Grasping
• Grasping a long object
A biologically Inspired Fitness
Function for Robot Grasping
• Grasping a long object with different orientation...
A biologically Inspired Fitness
Function for Robot Grasping
• Adjusting the fitness function itself to different kinds of
objects, raccoon style grasp
Conclusions
• An approach for guiding grasp selection choices for
multifingered robot hands based on GA and GP
• Ability of the GA and GP to arrive at sensible grasps
from an initial semi-random set of candidate grasps
• The approach can result in grasps which would not
have been initially selected by the user