Grasping Unknown Objects in Clutter by Superquadric Representation
This addresses robotic grasping in real-world clutter, but it is incremental as it builds on existing shape-based approaches.
The paper tackles the problem of grasping unknown objects in cluttered environments by using superquadric representations for partial views, achieving improved time-efficiency and accuracy compared to learning-based methods.
In this paper, a quick and efficient method is presented for grasping unknown objects in clutter. The grasping method relies on real-time superquadric (SQ) representation of partial view objects and incomplete object modelling, well suited for unknown symmetric objects in cluttered scenarios which is followed by optimized antipodal grasping. The incomplete object models are processed through a mirroring algorithm that assumes symmetry to first create an approximate complete model and then fit for SQ representation. The grasping algorithm is designed for maximum force balance and stability, taking advantage of the quick retrieval of dimension and surface curvature information from the SQ parameters. The pose of the SQs with respect to the direction of gravity is calculated and used together with the parameters of the SQs and specification of the gripper, to select the best direction of approach and contact points. The SQ fitting method has been tested on custom datasets containing objects in isolation as well as in clutter. The grasping algorithm is evaluated on a PR2 and real time results are presented. Initial results indicate that though the method is based on simplistic shape information, it outperforms other learning based grasping algorithms that also work in clutter in terms of time-efficiency and accuracy.