Adaptive Robust Kernels for Non-Linear Least Squares Problems

State estimation is a key ingredient in most robotic systems. Often, state performed using some form of least squares minimization. Basically, all error minimization procedures that work on real-world data use robust kernels as the standard way for dealing with outliers data. These kernels, however, are often hand-picked, sometimes different combinations, and their parameters need to be tuned manually particular problem. In this letter, we propose generalized kernel family, which automatically based distribution residuals includes common m-estimators. We tested our adaptive two popular problems robotics, namely ICP bundle adjustment. The experiments presented letter suggest approach provides higher robustness while avoiding manual tuning parameters.