Unsupervised Tuning of Filter Parameters Without Ground-Truth Applied to Aerial Robots

摘要

Autonomous robots heavily rely on well-tuned state estimation filters for successful control. This letter presents a novel automatic tuning strategy for learning filter parameters by minimizing the innovation, i.e., the discrepancy between expected and received signals from all sensors. The optimization process only requires the inputs and outputs of the filter without ground truth. Experiments were conducted with the Crazyflie quadrotor, and all parameters of the extended Kalman filter are well tuned after one 10-s manual flight. The proposed method has multiple advantages, of which we demonstrate two experimentally. First, the learned parameters are suitable for each individual drone, even if their particular sensors deviate from the standard, e.g., by being noisier. Second, this manner of self-tuning allows one to effortlessly expand filters when new sensors or better drone models become available. The learned parameters result in a better state estimation performance than the standard Crazyflie parameters.