Note: This checkpoint was trained jointly on four LIBERO sub-tasks. Starting from this checkpoint and further fine-tuning on a single task for 10k–20k steps typically yields an additional 3–5% performance improvement.