A three-stage framework from imitation to iterative offline and last-mile online RL that achieves near-perfect success with high efficiency and robustness.