Henry Lu

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2015-225

December 1, 2015

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-225.pdf

Learning from multiple demonstrations using non-rigid point cloud registration is an effective method for learning manipulation of deformable objects. Using non-rigid registration, we compute a warping function that transforms the demonstrated trajectory in each demonstration into the current scene. In the first section we present a technique for learning force-based manipulation from multiple demonstrations. Our method utilizes the variation between demonstrations to learn a variable-impedance control strategy that trades off force and position errors, providing the correct level of compliance that applies the necessary forces at different stages in the trajectory. The resulting force-augmented trajectory is effective in manipulating deformable objects that are variants of those used in demonstration, when the traditional kinematic method fails. In the second section we present a trajectory-aware non-rigid registration method that uses multiple demonstrations to focus the registration process on points that are more relevant to the task. This method allows us to handle significantly greater visual variation in the demonstrations than previous methods that are not trajectory-aware. When introducing irrelevant variations in the object geometry and random noise in the form distractor objects, trajectory-aware registration proves to be robust and capable of extracting the true goal of the task hidden in the similarities of the demonstrations. The naive single demonstration method and an ablated implementation of the trajectory-aware registration often fail in these situations. We evaluate our approaches on challenging tasks such as towel folding, knot tying, and object grasping from a box.

Advisors: Pieter Abbeel

---

BibTeX citation:

@mastersthesis{Lu:EECS-2015-225,
    Author= {Lu, Henry},
    Title= {Learning Manipulation of Deformable Objects from Multiple Demonstrations},
    School= {EECS Department, University of California, Berkeley},
    Year= {2015},
    Month= {Dec},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-225.html},
    Number= {UCB/EECS-2015-225},
    Abstract= {Learning from multiple demonstrations using non-rigid point cloud registration is an effective method for learning manipulation of deformable objects. Using non-rigid registration, we compute a warping function that transforms the demonstrated trajectory in each demonstration into the current scene. In the first section we present a technique for learning force-based manipulation from multiple demonstrations. Our method utilizes the variation between demonstrations to learn a variable-impedance control strategy that trades off force and position errors, providing the correct level of compliance that applies the necessary forces at different stages in the trajectory. The resulting force-augmented trajectory is effective in manipulating deformable objects that are variants of those used in demonstration, when the traditional kinematic method fails. In the second section we present a trajectory-aware non-rigid registration method that uses multiple demonstrations to focus the registration process on points that are more relevant to the task. This method allows us to handle significantly greater visual variation in the demonstrations than previous methods that are not trajectory-aware. When introducing irrelevant variations in the object geometry and random noise in the form distractor objects, trajectory-aware registration proves to be robust and capable of extracting the true goal of the task hidden in the similarities of the demonstrations. The naive single demonstration method and an ablated implementation of the trajectory-aware registration often fail in these situations. We evaluate our approaches on challenging tasks such as towel folding, knot tying, and object grasping from a box.},
}

EndNote citation:

%0 Thesis
%A Lu, Henry 
%T Learning Manipulation of Deformable Objects from Multiple Demonstrations
%I EECS Department, University of California, Berkeley
%D 2015
%8 December 1
%@ UCB/EECS-2015-225
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-225.html
%F Lu:EECS-2015-225