Graspit!: A Visualization and Simulation Engine
for Grasping Research

Peter Allen, Matei Ciocarlie and Corey Goldfeder
Department of Computer Science
Columbia University

Abstract

Robotic hands are still a long way from matching the grasping and
manipulation capability of their human counterparts.  One way to push
the research further along is to use computer modeling and simulation
to learn more about human and robotic grasping. We have created a
publicly available simulator called Graspit! to serve this purpose. It
can accommodate a wide variety of hand designs, and it can evaluate
grasps formed by these hands, as well as perform full dynamic
simulation of the grasping process. In this talk, we will describe a
wide range of research on intelligent grasping that we have performed
using the Graspit! simulator.

The first project is to use low-dimensional grasping subspaces to plan
stable grasps. The reduced dimensionality framework enables the
simulator to perform fast searches for effective hand postures and
thus achieve the computational rates needed for on-line interaction
with a human operator. We have applied this method for building an
on-line grasp planner that allows a human user to perform dexterous
grasping tasks using an artificial hand, even if the user has no
direct control over finger posture. The motivation for this project is
to enable the simulator to serve as an interface between the human and
automated components of a prosthetic system; we will discuss the
interplay between the two and present interesting new directions for
research in interactive grasping.  The second project is to
pre-compute stable grasps using GraspIt! on a wide variety of object
models to create an indexed database of models that can be efficiently
searched to find stable grasps on sensed objects in the environment.