Reinforcement Learning at RSL

Reinforcement learning is based on the concept of learning from rewards. If a robot is given rewards when its performance is good it can learn to improve its behaviour. The learning process can replace calibration and in the long-term could reduce programming effort and improve flexibility.

Many existing reinforcement learning algorithms treat continuous variables, like speed, by discretising them into a few levels (for example slow, medium, fast). Discretising does not allow smooth control.

The Robotic Systems Laboratory has developed an algorithm that allows reinforcement learning with continuous variables and tested the algorithm on real robots with real vision Systems.

Experiments

Learned visual servoing for a Nomad 200 mobile robot:
"Reinforcement Learning for Visual Servoing of a Mobile Robot," by C. Gaskett, L. Fletcher, and A. Zelinsky, in Proceedings of the Australian Conference on Robotics and Automation (ACRA2000), (Melbourne Australia, August 2000).

Learned wandering for a Nomad 200 mobile robot:
"Reinforcement Learning for a Vision Based Mobile Robot," by C. Gaskett, L. Fletcher, and A. Zelinsky, in Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS2000) © IEEE, (Takamatsu, Japan, October 2000).

Learned smooth fixation movement for the HyDrA active vision head (unpublished):

see also: A survey of continuous state and action Q-learning approaches:
"Q-Learning in Continuous State and Action Spaces," by C. Gaskett, D. Wettergreen, and A. Zelinsky, in Proceedings of the 12th Australian Joint Conference on Artificial Intelligence © Springer-Verlag, (Sydney, Australia, December 1999).

Other publications: Chris Gaskett's homepage