In my research I consider Robot Learning from Demonstration (RLfD) as
a means of performing Human-Robot Policy Transfer. The goal is to
allow a human user to instantiate, on a robot, an autonomous control
policy to perform a desired task without performing a detailed
analysis of the task itself or explicit coding. My work focuses on
using statistical machine learning approaches to infer these
controllers from noisy, suboptimal, and ambiguous human
In my dissertation, I developed an approach to extract an unknown
number of overlapping finite-state-machine states from unsegmented
demonstration. By assigning each of multiple 'correct' actions to
different subtasks, the robot can then execute only one instead of
averaging them, a potentially dangerous solution.
Currently, I am looking at learning from failed demonstrations by
modeling the demonstrations, and then deliberately avoiding them
during task reproduction. I am also investigating an alternate
parameterization of trajectories (as single high-dimensional points)
to take greater advantage of covariance information.
Much of my work deals with Gaussians, particularly Gaussian Mixture
Models and Gaussian Processes. You can read my most recent work-related thoughts on
AAAI 2011 learning from demonstration challenge.
AAAI 2011 Robot Challenge - LfD, San Francisco, CA, July 2011.
[ bib |
Getting robots to do what you want (even if you don't know what that is,
Willow Garage, Menlo Park, CA, June 2011.
[ bib |
Imitation and reinforcement learning from failed demonstrations.
Daniel H Grollman and Aude G Billard.
In ICML workshop on New Developments in Imitation Learning,
Bellevue, WA, June 2011.
[ bib |
EPFL, Ecole Polytechnique Federale de Lausanne
EPFL-STI-I2S-LASA, Station 9
CH 1015, Lausanne, Switzerland