EPFL, A university in Switzerland, seems to be at the forefront of innovative new technological developments, and this is no slouch either. From the schools Learning Algorithms and Systems Laboratory, Daniel Grollman and Aube Billard developed a system in which a robot will learn a task from a human and perform it until near perfect. For example, catapulting a ball into a basket/goal. A human will show it what to do, not even with high precision, and the robot will perform the task in a "repeat until convergence or success" way. A model is formed in the demonstration, and the bot refines the model through Dynamic Systems and Gaussian Mixture Models. The concept was originally to "[optimize] the learning behavior beyond the demonstrator's capabilities." Attached to this post is a whitepaper on the project with all the gritty detail.
Eavesdropper
Attachments:
| GrollmanBillard_ICRA2011.pdf |
