Incremental learning of robot dynamics using random features

Analytical models for robot dynamics often perform suboptimally in practice, due to various non-linearities and the difficulty of accurately estimating the dynamic parameters. Machine learning techniques are less sensitive to these problems and therefore an interesting alternative for modeling robot dynamics. We propose a learning method that combines a least squares algorithm with a non-linear feature mapping and an efficient update rule. Using data from five different robots, we show that the method can accurately model manipulator dynamics, either when trained in batch or incrementally. Furthermore, the update time and memory usage of the method are bounded, therefore allowing use in real-time control loops.