Towards Using Active Learning Methods for Human-Seat Interactions To Generate Realistic Occupant Motion

The paper presents a sound proof-of-concept active-learning-in-the-loop pipeline and explicitly states that convergence of either the OCP or the ML training is not guaranteed, while empirically showing a decreasing RMSE trend and a high-penetration underestimation bias. The model’s solution proposes a convergence proof via a Schauder fixed-point argument and RMSE subsequence convergence, but it hinges on strong additional assumptions (continuity of the OCP solution map in the contact law, an interpolating and equicontinuous training operator, compactness/equicontinuity of the surrogate class, and a memoryless FE oracle) that are neither stated nor justified by the paper’s setup. Hence, the paper lacks theory (by design) and the model’s proof does not apply to the described implementation without substantial new hypotheses.