Mazda Ahmadi, Peter Stone
Two main challenges of robot action planning in real domains are uncertain action effects and dynamic environments. In this paper, an instance-based action model is learned empirically by robots trying actions in the environment. Modeling the action planning problem as a Markov decision process, the action model is used to build the transition function. In static environments, standard value iteration techniques are used for computing the optimal policy. In dynamic environments, an algorithm is proposed for fast replanning, which updates a subset of state-action values computed for the static environment. As a test-bed, the goal scoring task in the RoboCup 4-legged scenario is used. The algorithms are validated in the problem of planning kicks for scoring goals in the presence of opponent robots. The experimental results both in simulation and on real robots show that the instance-based action model boosts performance over using parameterized models as done previously, and also replanning significantly improves over original off-line planning.
Subjects: 17. Robotics; 1.11 Planning
Submitted: Jun 6, 2006