Ph.D Student | Bunis Hallel |
---|---|

Subject | Caging-Based Grasping via Minimalistic Robot Hands |

Department | Department of Mechanical Engineering |

Supervisor | Professor Elon Rimon |

Full Thesis text |

Caging offers a robust method for grasping objects with multi-finger robot hands. To grasp an object, the robot hand is first configured in a cage formation around the object. The cage formation allows the object some freedom to move within the hand, but prevents it from escaping the hand. In the case of squeezing caging the hand is then closed, while in the case of stretching caging the hand is opened, until the object is completely immobilized by the fingers, without the need for a precise placement of the fingers relative to the object.

The thesis presents several novel caging-to-grasping algorithms for polygonal objects and robot hands that open and close according to a single parameter. The robot hands are comprised of three point or disc fingers, or two fingers that lock the objects against a linear wall. The configuration space of such hands is four-dimensional. Given a user specified target immobilizing grasp of an object, the algorithms compute the set of cage formations associated with it. For squeezing caging, starting from any cage formation in the set and closing the hand guarantees that the object will remain caged throughout the caging-to-grasping process, until an immobilizing grasp is reached. For stretching caging, opening the hand immobilizes the object.

Each algorithm in this thesis is
based on the construction of a *caging graph* embedded in the hand's two-dimensional
contact space. Contact space parameterizes all the hand configurations at which
at least two fingers contact the object. The nodes of the caging graph
represent frictionless equilibrium grasps of the object, as well as grasps
where a finger contacts a vertex of the polygonal object. The caging graph
edges represent feasible hand motions between hand configurations corresponding
to the caging graph nodes. Starting from the target immobilizing grasp, the set
of cage formations is determined by the critical cage formation that allows the
object to escape the hand for the first time. The critical cage formation forms
an equilibrium grasp of the object, and is hence represented by a graph node.
The difficult task of computing the caging set in the hand's four-dimensional
configuration space is thus reduced to a simple task of constructing and
searching the caging graph for the critical cage formation.

The first caging-to-grasping
algorithm presented in the thesis considers hands that squeeze the object in
order to immobilize it, while forming *equilateral triangle* finger
formations. Such hands are simple mechanically, however they cannot immobilize
every generic polygon. The second algorithm presented in the thesis extends the
first algorithm to hands that form any family of *similar triangle* finger
formations. This generalization increases the possible grasps that can be used
to immobilize an object. The third algorithm considers locking polygonal
objects *against a linear wall* using robot hands having two point or
disc-fingers. The wall is treated as a third finger of an equivalent
three-finger hand. The fourth algorithm considers *stretching grasps* by
three-finger robot hands that form a family of similar triangle formations, thus
expanding the grasp selection.