The evolution of superquadric modeling and grasping

[pattacini]

We recently extended the superquadric modeling and grasping pipeline to deal with more complex objects!

Our robots are now able to reconstruct 3D models and to compute grasp candidates for a large number of objects:

Technology

We use nonlinear constrained optimization both to fit the multi-superquadric model to the partial point cloud of the object and to obtain the grasping poses.

Perks

• The multi-superquadric model is the result of a splitting and merging procedure.
• When computing a grasp candidate for a specific portion of the object the other portions are considered as obstacles to avoid object penetration with the robot hand.

Outlook

• We're going to test the multi-superquadric approach on tool grasping with the iCub.
• We're planning to use our modeling and grasping strategy to guide RL training on our robots.

Code

A new and efficient implementation of our (single and multiple) superquadric modeling and grasping algorithms is now available as a library: https://github.com/robotology/superquadric-lib.

Publications

A new publication about multi-superquadric modeling and grasping is coming soon!

For now, you can have a look at the other papers explaining the evolution of our method:

• G. Vezzani, U. Pattacini, and L. Natale, A grasping approach based on superquadric models, ICRA 2017.
• G. Vezzani, U. Pattacini, G. Pasquale, and L. Natale, Improving Superquadric Modeling and Grasping with Prior on Object Shapes, ICRA 2018.

Credits go to @giuliavezzani

[jimingda]

Hi, what i want to do is to grasp the wheel by icub hands, so i find here, can i use this tool in my application?

[jimingda]

Maybe the grasp in red ball demo is more helpful to me. I will refer to that first.

[pattacini]

Given that you're equipped with a perfect model of the object of interest, I suggest employing it and carrying out ad-hoc inverse kinematics, specifically customized to provide grasp candidates for the steering wheel.

[jimingda]

Thanks@pattacini
Are there any tutorial of inverse kinematics or forward kinematics about fingers like thumb index and middle finger. Currently i have followed the air-hocky project and generated the trajectory of hand positions and poses on both arms, but still haven't control the fingers.

[pattacini]

Hi @jimingda

As far as I know, there's no specific tutorial for precision grasp. We did some work in the past on 3-fingers grasping¹, although it's definitely no longer maintained².

That said, I wouldn't go for precision-grasp in your context, but rather for power-grasp as you need to exert a non-negligible amount of force to steer the wheel.

A tutorial on power-grasp based on the cardinal points method can be found at https://github.com/robotology/icub-gazebo-grasping-sandbox.

Footnotes

1. Three-Finger Precision Grasp on Incomplete 3D Point Clouds. ↩

2. Code is still available at https://github.com/robotology/grasp anyway. ↩