Validation of Features for Characterizing Robot Grasps

This paper addresses the problem of characterizing robot grasps for unmodeled objects. We propose a set of intrinsic object features that are computed from the object image and the geometry of the robot hand. These features are validated by feeding them to neural networks which are trained with experimental data obtained with a humanoid robot. The results suggest that our features are actually suitable for predicting the reliability of a grip.

[1]  Eris Chinellato.  Robust Strategies for Selecting Vision-Based Planar Grasps of Unknown Objects with a Three-Finger Hand , 2002 .

[2]  Robert B. Fisher,et al.  Ranking planar grasp configurations for a three-finger hand , 2003, 2003 IEEE International Conference on Robotics and Automation (Cat. No.03CH37422).

[3]  Antonio Morales,et al.  Vision-based computation of three-finger grasps on unknown planar objects , 2002, IEEE/RSJ International Conference on Intelligent Robots and Systems.

[4]  Martin A. Riedmiller,et al.  A direct adaptive method for faster backpropagation learning: the RPROP algorithm , 1993, IEEE International Conference on Neural Networks.

[5]  Antonio Morales,et al.  An experiment in constraining vision-based finger contact selection with gripper geometry , 2002, IEEE/RSJ International Conference on Intelligent Robots and Systems.