Learning to Approach a Moving Ball with a Simulated Two-Wheeled Robot
We show how a two-wheeled robot can learn to approach a moving ball using Reinforcement Learning. The robot is controlled by setting the velocities of its two wheels. It has to reach the ball under certain conditions to be able to kick it towards a given target. In order to kick, the ball has to be in front of the robot. The robot also has to reach the ball at a certain angle in relation to the target, because the ball is always kicked in the direction from the center of the robot to the ball. The robot learns which velocity differences should be applied to the wheels: one of the wheels is set to the maximum velocity, the other one according to this difference. We apply a REINFORCE algorithm  in combination with some kind of extended Growing Neural Gas (GNG)  to learn these continuous actions. The resulting algorithm, called ReinforceGNG, is tested in a simulated environment with and without noise.
Unable to display preview. Download preview PDF.
- 2.Fritzke, B.: A growing neural gas network learns topologies. In: Tesauro, G., Touretzky, D., Leen, T. (eds.) Advances in Neural Information Processing Systems, vol. 7, Cambridge, MA (1995)Google Scholar
- 4.Riedmiller, M., Janusz, B.: Using neural reinforcement controllers in robotics. In: Proc. 8th Australian Conference on Artificial Intelligence, Canberra (1995)Google Scholar
- 5.Gaskett, C., Wettergreen, D., Zelinsky, A.: Q-learning in continuous state and action spaces. In: Australian Joint Conference on Artificial Intelligence, pp. 417–428 (1999)Google Scholar
- 6.Sutton, R., Barto, A.: Reinforcement Learning. An Introduction. MIT Press, Cambridge (2000)Google Scholar
- 7.Martinetz, T.: Competitive hebbian learning rule forms perfectly topology preserving maps. In: Proc. of ICANN 1993, pp. 427–434. Springer, Heidelberg (1993)Google Scholar