Robust humanoid vehicle ingress with regression and whole body control

Humanoid vehicle ingress-a prerequisite for autonomous driving-remains a critical unsolved challenge due to its sensitivity to initial conditions and unstructured environments. While model-based controllers fail under pose variations, Deep Reinforcement Learning (DRL) methods face computational bottlenecks and dependency on continuous localization. This paper proposes a hybrid control framework that bridges adaptability and efficiency by integrating whole-body control (WBC) and regression-driven adjustments. Our method trains regression models on DRL-generated simulation data to map initial pose deviations to future movement corrections, eliminating real-time DRL inference and dependency on continuous localization. In simulation, the controller achieves 97.6% ingress success with varying starting poses. Real-world experiments also validate that the proposed method enables humanoid vehicle ingress from different starting poses.