Matthew Matl

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2019-119

August 16, 2019

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-119.pdf

Robots in material-handling applications will need the ability to grasp and manipulate the huge variety of items that pass through warehouses every day. Handling this level of diversity requires the use of multiple end effectors, as each class of manipulator cannot handle certain types of objects. For example, parallel-jaw grippers fail on larger items, while vacuum-based end effectors cannot grasp porous or tiny objects. In this report, I present a summary of published research conducted by my colleagues and myself – specifically, Dex-Net 3.0 and Dex-Net 4.0 – that proposes methods for planning grasps for multiple classes of grippers on unknown objects. In Dex-Net 3.0, we propose a compliant suction contact model that geometrically com- putes the quality of the seal between a suction cup and a target surface and measures the ability of a suction grasp to resist an external gravity wrench. We then use this model to generate millions of suction grasp attempts in simulation and train a Grasp Quality Convo- lutional Neural Network (GQ-CNN) to predict grasp outcomes from synthetically-rendered point clouds. We evaluate the resulting system in 350 physical trials on an ABB YuMi fitted with a pneumatic suction gripper and achieve success rates of 98% on basic objects and 82% on more complex objects. In Dex-Net 4.0, we propose a unified contact model that brings the Dex-Net 3.0 suction model together with a point-contact model for parallel-jaw grippers. With this unified model, we generate a joint training dataset of over 5 million simulated grasps for both grippers and train GQ-CNNs for each gripper on that dataset. By selecting grasps from the GQ-CNN with the highest predicted confidence, we can decide between using a suction-cup end-effector and a parallel-jaw gripper when grasping objects from bins. In physical experiments, the Dex-Net 4.0 system consistently clears bins of up to 25 novel objects with a reliability greater than 95%. Code, datasets, the published papers, and supplemental material can be found at http://berkeleyautomation.github.io/dex-net.

Advisors: Ken Goldberg

---

BibTeX citation:

@mastersthesis{Matl:EECS-2019-119,
    Author= {Matl, Matthew},
    Editor= {Bajcsy, Ruzena and Goldberg, Ken},
    Title= {Learning Robotic Grasping Policies for Suction Cups and Parallel Jaws in Simulation},
    School= {EECS Department, University of California, Berkeley},
    Year= {2019},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-119.html},
    Number= {UCB/EECS-2019-119},
    Abstract= {Robots in material-handling applications will need the ability to grasp and manipulate the huge variety of items that pass through warehouses every day. Handling this level of diversity requires the use of multiple end effectors, as each class of manipulator cannot handle certain types of objects. For example, parallel-jaw grippers fail on larger items, while vacuum-based end effectors cannot grasp porous or tiny objects. In this report, I present a summary of published research conducted by my colleagues and myself – specifically, Dex-Net 3.0 and Dex-Net 4.0 – that proposes methods for planning grasps for multiple classes of grippers on unknown objects.
In Dex-Net 3.0, we propose a compliant suction contact model that geometrically com- putes the quality of the seal between a suction cup and a target surface and measures the ability of a suction grasp to resist an external gravity wrench. We then use this model to generate millions of suction grasp attempts in simulation and train a Grasp Quality Convo- lutional Neural Network (GQ-CNN) to predict grasp outcomes from synthetically-rendered point clouds. We evaluate the resulting system in 350 physical trials on an ABB YuMi fitted with a pneumatic suction gripper and achieve success rates of 98% on basic objects and 82% on more complex objects.
In Dex-Net 4.0, we propose a unified contact model that brings the Dex-Net 3.0 suction model together with a point-contact model for parallel-jaw grippers. With this unified model, we generate a joint training dataset of over 5 million simulated grasps for both grippers and train GQ-CNNs for each gripper on that dataset. By selecting grasps from the GQ-CNN with the highest predicted confidence, we can decide between using a suction-cup end-effector and a parallel-jaw gripper when grasping objects from bins. In physical experiments, the Dex-Net 4.0 system consistently clears bins of up to 25 novel objects with a reliability greater than 95%.
Code, datasets, the published papers, and supplemental material can be found at http://berkeleyautomation.github.io/dex-net.},
}

EndNote citation:

%0 Thesis
%A Matl, Matthew 
%E Bajcsy, Ruzena 
%E Goldberg, Ken 
%T Learning Robotic Grasping Policies for Suction Cups and Parallel Jaws in Simulation
%I EECS Department, University of California, Berkeley
%D 2019
%8 August 16
%@ UCB/EECS-2019-119
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-119.html
%F Matl:EECS-2019-119