Towards Robust State Estimation by Boosting the Maximum Correntropy Criterion Kalman Filter with Adaptive Behaviors

Robust state estimation remains as an open prob- lem in robotics, specially when the robot is meant to navigate in perception degraded environments. As a step towards this robustness, in this paper we propose a new sensor fusion approach that extends the Maximum Correntropy Criterion Kalman Filter (MCCKF) by incorporating adaptive behaviors. The MCCKF outperforms traditional Kalman filters in the presence of large outliers or sensor failures. However, MCCKFs have a clear dependency on the size of their Kernel Bandwith (KB), which is a user-selected parameter and has a big impact on the resulting filter estimation. Further, MCCKFs have fixed system and observation noise covariance matrices, which usually require tedious tuning efforts and is a common source of divergence. To overcome these issues, we transform the MCCKF to an adaptive-MCCKF (AMCCKF) that dynamically modifies the KB size and both system and observation noise covariance matrices. The validation of the fusion algorithm is shown through real experiments using a small aerial platform and a ground robot in a loosely-coupled sensor fusion architecture. The presented approach is part of the solution used by the COSTAR team participating at the DARPA Subterranean Challenge.