Samantha Wathugala

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2019-45

May 15, 2019

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-45.pdf

Navigation is often required in order to perform complex grasping tasks such as cleaning unstructured household rooms. Although there is extensive prior work in applying reinforcement learning to grasping and navigation, these tasks have largely been separate from each other. Our work's primary contribution is to determine whether learning to navigate in a way that might improve grasping would provide an increase of grasp success.

Exploring this objective and answering this question led to two stages of work, both grounded in the complex task of cleaning up toys scattered in a room. In the first stage, assuming that we could navigate to any pre-grasp orientation, we investigated whether certain orientations were more amenable to successful grasps. We trained probability of success models as a grasping policy, to potentially introduce a heuristic that we could judge graspability off of. We found that, though that heuristic was good enough to produce reasonably effective grasping policies, it was not good enough to be used in lieu of actual grasp success. We then ran those policies on the physical robot with a single toy and observed the actual grasp success rates and images collected. We did find that for each grasping policy, there existed toy orientations that proved to be more difficult to grasp than others.

With this idea validated, we next implement a navigation policy to approach toys at more graspable orientations, and compare its grasp success to the original navigation policy. We use a simple nearest neighbor's approach to determine which pre-grasp orientations of a certain toy were good. In real time, before each potential grasp, we use a Structure-from-Motion model and the robot camera's current image to estimate the angle of the robot's approach to this toy, and then determine whether to grasp or to continue to circle the toy based on how successful past grasps at that approach have been. From implementing and running this policy, we find that this navigation policy indeed improves grasp success from 0.59 to 0.80. This demonstrates the hypothesis that was previously only validated: training a navigation policy based on former grasping experience can and does improve grasp success in tasks that involve both.

Advisors: Pieter Abbeel

---

BibTeX citation:

@mastersthesis{Wathugala:EECS-2019-45,
    Author= {Wathugala, Samantha},
    Title= {Leveraging Mobility for Robot Grasping Tasks},
    School= {EECS Department, University of California, Berkeley},
    Year= {2019},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-45.html},
    Number= {UCB/EECS-2019-45},
    Abstract= {Navigation is often required in order to perform complex grasping tasks such as cleaning unstructured household rooms. Although there is extensive prior work in applying reinforcement learning to grasping and navigation, these tasks have largely been separate from each other. Our work's primary contribution is to determine whether learning to navigate in a way that might improve grasping would provide an increase of grasp success.

Exploring this objective and answering this question led to two stages of work, both grounded in the complex task of cleaning up toys scattered in a room. In the first stage, assuming that we could navigate to any pre-grasp orientation, we investigated whether certain orientations were more amenable to successful grasps. We trained probability of success models as a grasping policy, to potentially introduce a heuristic that we could judge graspability off of. We found that, though that heuristic was good enough to produce reasonably effective grasping policies, it was not good enough to be used in lieu of actual grasp success. We then ran those policies on the physical robot with a single toy and observed the actual grasp success rates and images collected. We did find that for each grasping policy, there existed toy orientations that proved to be more difficult to grasp than others.

With this idea validated, we next implement a navigation policy to approach toys at more graspable orientations, and compare its grasp success to the original navigation policy. We use a simple nearest neighbor's approach to determine which pre-grasp orientations of a certain toy were good. In real time, before each potential grasp, we use a Structure-from-Motion model and the robot camera's current image to estimate the angle of the robot's approach to this toy, and then determine whether to grasp or to continue to circle the toy based on how successful past grasps at that approach have been. From implementing and running this policy, we find that this navigation policy indeed improves grasp success from 0.59 to 0.80. This demonstrates the hypothesis that was previously only validated: training a navigation policy based on former grasping experience can and does improve grasp success in tasks that involve both.},
}

EndNote citation:

%0 Thesis
%A Wathugala, Samantha 
%T Leveraging Mobility for Robot Grasping Tasks
%I EECS Department, University of California, Berkeley
%D 2019
%8 May 15
%@ UCB/EECS-2019-45
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-45.html
%F Wathugala:EECS-2019-45