Robotic Manipulation with a Human in the Loop - The Pipeline for Learning by Demonstration

Sunil Srinivasan, Sebastian Schweigert, Jiewen Sun, James Su and Mark Jouppi

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2015-87
May 14, 2015

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-87.pdf

Current practical-use robots are designed for very specific tasks and cannot be generalized to arbitrary tasks. However, to spread the use of generalized-task robots, they must be able to perform precise tasks, but must also be cost-effective; the high precision that single-task robots usually have is accompanied by a high hardware cost. We propose a system that allows lower-cost robots to learn and attempt tasks that require high precision while failing safely, allowing the system to detecting portions of the task that it cannot complete alone, and then requesting human assistance for precision or accuracy that it cannot achieve. The system generates a model of the task from demonstrations, then analyzes the model to determine the accuracy required during different segments of the task. Subsequent comparison to the robot's known tolerance specifications allows the system to flag portions the robot cannot handle. Our approach generates the task model from a set of processed trajectories and a Gaussian mixture model, then extracts an ideal trajectory for the task as well as a set of standard deviations for each dimension of the trajectory data, allowing us to parametrize the accuracy for each segment of the task. We also analyze industry factors to determine where such a system would fit in the current market for robotic technology, for a more comprehensive picture of where the technology stands.

Advisor: Ruzena Bajcsy


BibTeX citation:

@mastersthesis{Srinivasan:EECS-2015-87,
    Author = {Srinivasan, Sunil and Schweigert, Sebastian and Sun, Jiewen and Su, James and Jouppi, Mark},
    Title = {Robotic Manipulation with a Human in the Loop - The Pipeline for Learning by Demonstration},
    School = {EECS Department, University of California, Berkeley},
    Year = {2015},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-87.html},
    Number = {UCB/EECS-2015-87},
    Abstract = {Current practical-use robots are designed for very specific tasks and cannot be generalized to arbitrary tasks.  However, to spread the use of generalized-task robots, they must be able to perform precise tasks, but must also be cost-effective; the high precision that single-task robots usually have is accompanied by a high hardware cost.
We propose a system that allows lower-cost robots to learn and attempt tasks that require high precision while failing safely, allowing the system to detecting portions of the task that it cannot complete alone, and then requesting human assistance for precision or accuracy that it cannot achieve.  The system generates a model of the task from demonstrations, then analyzes the model to determine the accuracy required during different segments of the task.  Subsequent comparison to the robot's known tolerance specifications allows the system to flag portions the robot cannot handle.  Our approach generates the task model from a set of processed trajectories and a Gaussian mixture model, then extracts an ideal trajectory for the task as well as a set of standard deviations for each dimension of the trajectory data, allowing us to parametrize the accuracy for each segment of the task.
We also analyze industry factors to determine where such a system would fit in the current market for robotic technology, for a more comprehensive picture of where the technology stands.}
}

EndNote citation:

%0 Thesis
%A Srinivasan, Sunil
%A Schweigert, Sebastian
%A Sun, Jiewen
%A Su, James
%A Jouppi, Mark
%T Robotic Manipulation with a Human in the Loop - The Pipeline for Learning by Demonstration
%I EECS Department, University of California, Berkeley
%D 2015
%8 May 14
%@ UCB/EECS-2015-87
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-87.html
%F Srinivasan:EECS-2015-87