Vision‑language‑action (VLA) models have shown great promise for robotic manipulation, yet their post‑training still relies heavily on trajectory‑level supervised imitation of robot demonstrations. While this approach learns feasible actions, it forces policies to reproduce expert actions at every step, biasing them toward demonstration‑specific paths and offering little guidance for recovering to task‑progress states when execution deviates from the demonstrated trajectories. To overcome this limitation, we introduce OGPO—an Offline Goal‑conditioned Policy Optimization framework designed for recoverable VLA post‑training. Instead of treating offline demonstrations merely as step‑wise action labels, OGPO adaptively relabels trajectories with semantic key states and constructs goal‑conditioned process rewards to optimize the policy for reaching these task‑progress states. In this way, OGPO shifts VLA post‑training from trajectory‑level action imitation to key‑state reachability learning, enabling policies to robustly achieve goal states from feasible off‑demonstration states without additional online interaction. Experiments on LIBERO and MetaWorld demonstrate that OGPO consistently improves VLA policy performance over standard supervised fine‑tuning. Further evaluations on LIBERO‑Plus and the newly proposed mid‑execution perturbation benchmark LIBERO‑ReAct show that OGPO attains stronger robustness and goal‑directed recoverability under distribution shifts and recoverable off‑demonstration states.
Review