Task Parameteris(/z?)ed Learning

I've put together this short demo video for a talk I'm giving soon showing some recent work on task parameterised (TP) learning methods on our Sawyer Robot https://www.youtube.com/watch?v=i0dcedYD-II

Specifically, this is using a TP Gaussian Mixture model to encode the demonstrations, and then a Gaussian Mixture Regression to generate the new trajectories from the model.

This is a method developed by Sylvain Calinon et al and there's a nice software library to help get you started :) http://www.idiap.ch/software/pbdlib/

For a more detailed insight to TP methods, I highly recommend this paper http://calinon.ch/paper4018.htm

In order to incorporate grasping to the model, I parameterised the control of the AR10 hand and recorded the control signals as a continuous variable, which happily seems to work perfectly.

Starting to get the hang of this robot, looking forward to trying out more things.