Christopher Correa

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2019-80

May 17, 2019

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-80.pdf

Robots are expected to grasp complex 3D objects in a wide variety of situations. This task can be difficult when the object's pose prevents the robot from perceiving or executing grasps on the object. When robust grasps are not accessible, robots can execute non-prehensile actions such as pushing and toppling to change an object's 3D pose to provide access to robust grasps. We develop two planar pushing policies and evaluate each policy's ability to increase access to robust grasps for both parallel jaw grippers and vacuum suction grippers. Using an ABB YuMi arm, we execute each pushing policy on the same 1000 simulated and physical scenarios in which the quality of all accessible grasps is low, and measure the predicted grasp reliability before and after the push. These experiments suggest that pushing can be used effectively to expose robust parallel jaw grasps, but are less effective in exposing robust vacuum suction grasps. As a result, we explore using toppling to reveal flat object faces for vacuum suction grippers. We present a toppling model which characterizes the robustness of toppling a 3D object specified by a triangular mesh, using Monte Carlo sampling to account for uncertainty in object apose, friction coefficients, and push direction. We run 700 physical toppling experiments using the ABB Yumi arm to compare the performance of the proposed model against empirical outcomes. We find that the toppling model outperforms a baseline model by an absolute 26.9\% when comparing the total variation distance between each model's predicted probability distribution and the empirical distribution. We use the robust model as the state transition function in a Markov Decision Process (MDP) to plan optimal sequences of toppling actions to expose access to robust suction grasps. Data from 20,000 simulated experiments suggests the toppling policy can increase suction grasp reliability by 33.6\%.

Advisors: Ruzena Bajcsy

---

BibTeX citation:

@mastersthesis{Correa:EECS-2019-80,
    Author= {Correa, Christopher},
    Title= {Facilitating Robotic Grasping using Pushing and Toppling},
    School= {EECS Department, University of California, Berkeley},
    Year= {2019},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-80.html},
    Number= {UCB/EECS-2019-80},
    Abstract= {Robots are expected to grasp complex 3D objects in a wide variety of situations.  This task can be difficult when the object's pose prevents the robot from perceiving or executing grasps on the object.  When robust grasps are not accessible, robots can execute non-prehensile actions such as pushing and toppling to change an object's 3D pose to provide access to robust grasps.  We develop two planar pushing policies and evaluate each policy's ability to increase access to robust grasps for both parallel jaw grippers and vacuum suction grippers.  Using an ABB YuMi arm, we execute each pushing policy on the same 1000 simulated and physical scenarios in which the quality of all accessible grasps is low, and measure the predicted grasp reliability before and after the push.  These experiments suggest that pushing can be used effectively to expose robust parallel jaw grasps, but are less effective in exposing robust vacuum suction grasps.  As a result, we explore using toppling to reveal flat object faces for vacuum suction grippers.  We present a toppling model which characterizes the robustness of toppling a 3D object specified by a triangular mesh, using Monte Carlo sampling to account for uncertainty in object apose, friction coefficients, and push direction.  We run 700 physical toppling experiments using the ABB Yumi arm to compare the performance of the proposed model against empirical outcomes.  We find that the toppling model outperforms a baseline model by an absolute 26.9\% when comparing the total variation distance between each model's predicted probability distribution and the empirical distribution.  We use the robust model as the state transition function in a Markov Decision Process (MDP) to plan optimal sequences of toppling actions to expose access to robust suction grasps. Data from 20,000 simulated experiments suggests the toppling policy can increase suction grasp reliability by 33.6\%.},
}

EndNote citation:

%0 Thesis
%A Correa, Christopher 
%T Facilitating Robotic Grasping using Pushing and Toppling
%I EECS Department, University of California, Berkeley
%D 2019
%8 May 17
%@ UCB/EECS-2019-80
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-80.html
%F Correa:EECS-2019-80