Dextrous robot hands have many degrees of freedom. This enables the manipulation of
objects between the digits of the dextrous hand but makes grasp planning substantially
more complex than for parallel jaw grippers. Much of the work that addresses grasp
planning for dextrous hands concentrates on the selection of contact sites to optimise
stability criteria and ignores the kinematics of the hand. In more complete systems,
the paradigm of preshaping has emerged as dominant. However, the criteria for the
formation and placement of the preshapes have not been adequately examined, and
the usefulness of the systems is therefore limited to grasping simple objects for which
preshapes can be formed using coarse heuristics.
In this thesis a grasp metric based on stability and kinematic feasibility is introduced.
The preshaping paradigm is extended to include consideration of the trajectories that
the digits take during closure from preshape to final grasp. The resulting grasp family
is dependent upon task requirements and is designed for a set of "ideal" object-hand
configurations. The grasp family couples the degrees of freedom of the dextrous hand
in an anthropomorphic manner; the resulting reduction in freedom makes the grasp
planning less complex. Grasp families are fitted to real objects by optimisation of the
grasp metric; this corresponds to fitting the real object-hand configuration as close to
the ideal as possible. First, the preshape aperture, which defines the positions of the
fingertips in the preshape, is found by optimisation of an approximation to the grasp
metric (which makes simplifying assumptions about the digit trajectories and hand
kinematics). Second, the full preshape kinematics and digit closure trajectories are
calculated to optimise the full grasp metric.
Grasps are planned on object models built from laser striper range data from two
viewpoints. A surface description of the object is used to prune the space of possible
contact sites and to allow the accurate estimation of normals, which is required by the
grasp metric to estimate the amount of friction required. A voxel description, built by
ray-casting, is used to check for collisions between the object and the robot hand using
an approximation to the Euclidean distance transform.
Results are shown in simulation for a 3-digit hand model, designed to be like a simplified
human hand in terms of its size and functionality. There are clear extensions of the
method to any dextrous hand with a single thumb opposing multiple fingers and several
different hand models that could be used are described. Grasps are planned on a wide
variety of curved and polyhedral objects