Currently, KiRo disposes of only few playing skills. On the real
table it is very difficult to implement skills like stopping or
passing the ball because of noisy and incomplete sensor data and
the uncertainty in the actions taken. However, a simulator can
simplifiy the problem by providing more accurate and reliable
sensing and acting. This project aims at implementing skills for
stopping, passing and dribbling the ball in the simulted table soccer
game. However, the hand crafted skills should also be evaluated and
optimzed for the real table. Finally, already existing reinforcement
learning software should be extended for learning the skills.
Part A -
Hand-Crafted Skills for the Simulator
Approach:
Implement StopBall, DribbleBall and PassBall using the
KiRo-Simulator. Implement the following simple action selection for
two rods: the first rod stops the ball, dribbles it for a short while, passes it to the
second rod, which stops, dribbles and passes back to the first rod
which stops, dribbles...
Evaluation:
Document impressions of observed performance. Measure time
(averaged over several runs) until ball is not controlled
by neither one of the two playing rods anymore.
Delivery:
08.01.04
Part B -
Learned Skills
Approach:
Learn StopBall, DribbleBall and PassBall using and enhancing the
available reinforcement learning software. For getting familiar with
the software learn BlockBall first.
Evaluation:
Document impressions of observed performance. Measure time
(averaged over several runs) until ball is not controlled
by neither one of the two playing rods anymore. Compare each skill
an the overall behavior with the hand-cratfted skills.
Delivery:
22.01.04
Part C -
Skills on the Real Table
Approach:
Test the hand-crafted and the learned skills on the real
table. Adapt and optimzie the hand crafted skills for the use on the real table.
Evaluation:
Document impressions of observed performance. Describe
problems encountered on the real table.
Institut für
Informatik, Universität
Freiburg