The aim of DARPA bootstrap learning project is to create an electronic student that can learn a variety of tasks by taking lessons from a domain knowledgeable teacher. The student is attempting to mimic human learning and therefore must be capable of learning in wide variety of manners, some examples include `Learning by telling', 'Learning by example' and `Learning by feedback'. At UT we are focused on `Learning by feedback' and in particular applying reinforcement learning to a teacher-student setup.

Currently the DARPA page (http://bootstraplearning.com/) contains a very brief introduction but more information should be presented here at some stage. Other details can be found by looking at various press releases [1] [2] that were released in 2007.

Note: This bootstrap learning project should not be confused with Ben Kuipers and his students use of the term `bootstrap learning', this is an unfortunate naming conflict. However you should check out their work as its cool stuff !

I've been working on robotics research inside the RoboCup domian since 2002. In particular in the Four-Legged league using the Sony Aibo robot. From 2008 we are also using the Aldebaran Nao humanoid robot.

RoboCup forces you to build a complete multi-agent robotic system. To produce a successful robotic soccer agent you need to have all the key subsystems working well. These systems are Vision, Localisation, Behaviour and Motion. In general robotics research a lot of research is focused on building these sub-systems, however many researches tend to underestimate the effort (and skill) required to make all these systems work well together.

Over the years I have written code in all these major subsystems, but I tend to enjoy working in the behaviour and motion layers. In particular the area in which these two overlap, which I refer to as the robots `low level skills', such as dribbling, walking, kicking etc. These are the skills that can make-or-break your robot as they directly influence what high level decision options are available. It matters little if your robot has fantastic high level decision making code if it does not have the skills required to execute it !