Learning Autonomous Grasping Strategies for a Care Robot

Abstract

Autonomous grasping is a key requisite for the autonomy of robots.
However, grasping of unknown objects in domestic environments is difficult due to the presence of unpredictability and clutter.
In this paper, a novel algorithm capable of finding an unobstructed grasping pose on unknown regular object shapes in cluttered environments is proposed. At first, an axis aligned bounding box is fitted around the target object point cluster, proposing a limited set of possible grasps.
A Support Vector Machine is trained to take the decision of the preferred grasp via supervised Machine Learning. Such decision is based on the distribution of obstacles around the target object, inferred by sampling the OctoMap of the scene.
Four different versions of the algorithm have been trained, tested and validated with simulation and real data, with known and unknown target object shapes, in more than 120 different scenarios. The best performing algorithm version returned an unobstructed grasping position in more than 90\% of the cases.
Thanks to the the use of the OctoMap representation, significant results are achieved with a limited set of training examples, with a 10 to 10.000 fold reduction of training examples amount with respect to comparable state-of-the-art algorithms. The algorithm also proved itself capable to generalise to novel object shapes and from simulation to real data.