Robot-Control Based on Extended Markov Tracking: Initial Experiments

Extended Markov Tracking (EMT) is a computationally tractable method for the online estimation of Markovian system dynamics. In this paper, we present our initial experimentation with EMT-based control applied to robotic motion. In our experiments, a robot uses a predetermined mapping of the world onto an abstract model, over which EMT Control is applied; this dictates the choice of an abstract action, which in turn is mapped back into actual robot operation. Simulations in which a robot was constrained to follow a target show that although the abstract model was (intentionally) only weakly coherent with the real dynamics of the robot’s world, EMT Control was able to provide reasonable performance. We also demonstrate that EMT-provided data provides sensible information for action model calibration. We do so by constructing a calibration scheme based on a training technique and simple data statistics. The scheme is then validated by carrying out additional robot motion control simulations, using the calibrated abstract model.