Learning tactic-based motion models with fast particle smoothing

Abstract

Abstract-Learning parameters of a motion model is an important challenge for autonomous robots. We address the particular instance of parameter learning when tracking motions with a switching state-space model. We present a general algorithm for dealing simultaneously with both unknown fixed model parameters and state variables. Using an ExpectationMaximization approach, we apply a tactic-based multi-model particle filter to estimate the state variables in the E-step, and use particle smoothing to update the parameters in the M-step. We test our algorithm both in simulation and in a team robot soccer environment, as a substrate for applying the learned models to object tracking in a team. One of the soccer robots learns the actuation model of its teammate. The experimental results show that the particle smoothing efficiency is substantially increased and the tracking performance is significantly improved using the learned teammate actuation model. This paper addresses estimating state and learning motion models in such a hybrid-state system. We are interested in tracking the ball in a robot soccer domain. This is a highly dynamic and multi-agent environment. All the robots in the field can actuate over the ball, e.g., grab and kick the ball, making the motion model of the ball very complex [3]. The good news is that we can acquire information about the ball motion from multiple sources besides the sensor. First, the robot's tactics provides valuable information and a tacticbased motion modeling and tracking algorithm is introduced in such scenarios [4]. Second, when the robot is playing a game as a member of a team, the team coordination knowledge provides further information that can be incorporated into the motion modeling and tracking process. We based our work upon a plan-dependent tracking algorithm called play-based tracking [5]. I. INTRODUCTIONAny model consists of one or multiple parameters. Usually the model parameters are set by a human expert, based upon the experience with the environment and the robot. In this paper, we present a novel method of automating the procedure of acquiring this probabilistic motion model. This approach deals simultaneously with both unknown fixed