The pyramid wavefront sensor is an asset for an AO system thanks to its sensitivity. However, because it is a nonlinear sensor it comes with operational challenges. A convolutional method and a gain sensing camera allow to track the optical gains, which encode the sensitivity variations due to the nonlinearities. Tracking and compensating the optical gains is necessary to perform extreme adaptive optics and to operate the pyramid off-zero to compensate for the NCPA.This study focuses on the reliability of this method. A numerical twin of the bench PAPYRUS, developed for this study, shows a improvement of the performance by a factor 2.7 on the Strehl Ratio when compensating for the optical gains. The convolutional method is implemented for the PAPYRUS bench, allowing the first on-sky tracking of optical gains. The next main steps are to compensate for the optical gains in real-time, then to offset the pyramid in order to optimise fiber-injection, to compensate for NCPA and to provide AO generated dark hole for high-contrast imaging.