Online Stochastic Model Identification for Safe and Accurate Two-Wheeled Robot Control

Stochastic disturbances are inevitable for realworld robot systems, so for safe and efficient robot control, a stochastic robot motion model is required. However, it is not enough to simply estimate the stochastic robot motion model because underestimating the probability variation of control increases the risk of robot collisions. Conversely, overestimating the probability variation of control can decrease this risk, but the controller becomes conservative and less efficient. Therefore, in this paper, we propose a safe and accurate online estimation method for the diffusion term in the stochastic differential equation of a two-wheeled mobile robot. The proposed method utilizes model uncertainty to estimate a reasonably conservative model in the early stages of learning and then gradually improves the efficiency. Simulations and real-world experiments show that the proposed method can achieve higher estimation accuracy than other methods while keeping the underestimation of the robot's motion disturbance consistently at low level and gradually approaching the optimal solution as the training progresses.