Model-based sideslip estimation relies on knowing the tires' cornering stiffness, which changes as tires wear or heat up. Online adaptation of the stiffness can keep the estimator valid, but the abstract describes a conflict: this adaptation competes with a learned model correction that operates on the same filter innovation. If the learned corrector has enough parameters, it can absorb the innovation that should be driving the stiffness update.
The paper, available on arXiv, introduces "sensitivity-orthogonal residual learning" to address this competition. The idea, as the title suggests, is to shape the learned residual so that it is orthogonal to the sensitivity of the stiffness adaptation—preventing the corrector from cancelling the information needed for the stiffness estimate.
The abstract is short, so the full method and its experimental validation are not detailed here. What is clear from the source is the problem: a naive combination of learned residual correction and adaptive stiffness estimation can lead to the learned model swallowing the signal that should be used for adaptation. The proposed approach is a targeted solution to that identifiability issue.