Privacy-Preserving Grasp Planning

This project is maintained by BerkeleyAutomation

Privacy-Preserving Grasp Planning

Privacy-Preserving Grasping

Overview of privacy-preserving grasp planning. Industrial users label proprietary zones on the part with a graphical interface. The masked object is then transmitted to a Cloud-based grasp planner along with its stable poses. The grasp planner computes a set of grasps for each stable pose and returns the grasp sets to the user.

Abstract

To support industrial automation, systems such as GraspIt! and Dex-Net 1.0 provide ``Grasp Planning as a Service" (GPaaS). This can allow a manufacturer setting up an automated assembly line for a new product to upload part geometry via the Internet to the service and receive a ranked set of robust grasp configurations. As industrial users may be reluctant to share proprietary details of product geometry with any outside parties, this paper proposes a privacy-preserving approach and presents an algorithm where a masked version of the part boundary is uploaded, allowing proprietary aspects of the part geometry to remain confidential. One challenge is the tradeoff between grasp coverage and privacy: balancing the desire for a rich set of alternative grasps based on analysis of graspable surfaces (coverage) against the user's desire to maximize privacy. We introduce a grasp coverage metric based on dispersion from motion planning, and plot its relationship with privacy (the amount of the object surface that is masked). We implement our algorithm for Dex-Net 1.0 and present case studies of the privacy-coverage tradeoff on a set of 23 industrial parts. Results suggest that masking the part using the convex hull of the proprietary zone can provide grasp coverage with minor distortion to the object similarity metric used to accelerate grasp planning in Dex-Net 1.0.

References

Contributors

This is an ongoing project at UC Berkeley with contributions from:
Jeffrey Mahler, Brian Hou, Sherdil Niyaz, Florian T. Pokorny, Ken Goldberg
The project is supported in part by grants from Siemens.

Support or Contact

Please Contact Jeffrey Mahler jmahler@berkeley.edu or Prof. Ken Goldberg, Director of Automation Sciences Lab at goldberg@berkeley.edu

Creative Commons License
Privacy-Preserving Grasping by Automation Lab is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.