ARUP; ISBN: 978-0-9562121-5-3 - CMBBE 2012 - Cardiff University

predicting hand postures. In some cases, artificial intelligence techniques are used for
learning the hand inverse kinematics given the fingertips’ positions, as in [4].
From the robotics’ viewpoint, the questions posed by grasping are slightly different. For
a given object, grasp synthesis must provide the most appropriate set of contact points
and hand posture for grasping. This problem has been studied using analytical
approaches that have been nearly ruled out due to their poor results when it comes to
real implementations [6]. Currently, artificial intelligence algorithms are used, as those
based on artificial neural networks (ANN) [10]. These sensor-based approaches learn
the underlying rules of robot grasping without the need of explicit kinematic models, by
means of exploration. For instance, an ANN-based strategy was developed in [11] for
grasping tasks with a 5 degrees-of-freedom (dof) gripper. Similarly, a grasping model
was design in [13] for grasp synthesis of a 7-dof planar hand for circular and rectangular
objects. ANN methods have also been used in feasible contact point determination [1,
13]. In [13] an ANN able to learn generic grasping functions for simple 2-finger
grippers was developed. Still, it was not extended to multi-fingered hands due to the
lack of grasping information for an object (contact points). In [2] this problem was
tackled defining hand contact configurations, associated to feasible object contact zones.
Many researchers state that advances in the field of robot grasping and manipulation
require a better knowledge of human grasping [14]. This is especially true in the field of
service robotics, where robots move in human environments and interact with daily-life
objects [15]. The variety of robotics grasps studied so far is very scarce and depends on
the features of the particular robot hand. To deal with new hands or objects, some works
begin to use human grasping information to guide the robot grasp synthesis, as in [1],
where a data glove is used for training a neural network that produces robot grasping
postures from previously computed contact zones for different objects.
Along these lines, this paper aims at predicting the grasping posture for the index and
thumb fingers of human hands from characteristic hand data, the features of the object
to be grasped and the task to perform. Different feed-forward networks have been tested
for automatically providing hand postures given such inputs. This is the first step in the
elimination of the initial data collection phase required in a grasping study. The outputs
of the ANN are the fingers joint angles for the grasping posture. The ANN has been
trained with data collected from grasping experiments with daily-life objects (bottles)
and tasks (moving and pouring). Once trained, the network is able to predict grasping
postures for objects different to those used in training with an acceptable error.
ROLL MCP1
MCP2
ABD1
a) b)
Fig. 1. a) Bottles B1, B2, B3, and B4 (left to right); b) Cyberglove joint sensors.
DIP1
PIP2