“Attempts to fine-tune a sim-trained policy with real-world data at deployment time were unsuccessful because data collection was too slow and the process often broke the policy's behavior.”