EFC: Explicit Feedforward Compensation for Humanoid Standing Balance on a Simulated Maritime Platform

Stable standing on moving decks is a prerequisite for humanoid robots to perform inspection and intervention in maritime environments. A key challenge is to translate anticipated support surface motion into coordinated joint corrections while retaining learned whole-body feedback. We propose explicit feedforward compensation (EFC), which maps a filtered causal second-order roll forecast to bounded differential knee position offsets and combines this anticipatory channel with recurrent proximal policy optimization feedback for the Unitree H1-2 in MuJoCo. Across three independent training runs and matched 60 s evaluations under prescribed periodic roll–pitch excitation at 5∘, EFC exhibits 9.0% and 7.3% lower mean attitude RMSE than Prediction Input Only and Current Roll Knee Residual, respectively, over the recorded survival intervals, together with a 60 s restricted mean survival time of 56.20 s compared with 52.58 and 52.72 s. These results indicate that structured predictive knee compensation can complement recurrent feedback to improve posture regulation and sustained standing under the tested periodic deck conditions.