Sim2Act: Robust Simulation-to-Decision Learning via Adversarial Calibration and Group-relative Perturbation | ResearchPod