Learning Using Behavioral Primitives

Chris Atkeson

Abstract

We are exploring how to improve robot learning by using larger scale units of behavior (which we call "primitives", and others have called schemas, macros, and behaviors). We are using a tilt maze task and humanoid air hockey as experimental testbeds. We have implemented both learning from observation (watching a human do the tasks) and learning from practice. This talk will discuss memory-based approaches to primitive selection and subgoal and action generation.

 

Speaker Appointments

For appointments, please contact Chris Atkeson (cga@cmu.edu).