Odyssey has introduced PROWL, a reinforcement learning-driven framework designed to enhance world model training. The PROWL system uses a behavior-constrained RL agent to identify and address failure trajectories in world models, focusing on geometric, motion, visual consistency, and action responsiveness. This approach establishes a scalable feedback loop, utilizing a Prioritized Adversarial Trajectory buffer to prioritize unresolved challenges, thereby fostering continuous model improvement. The framework was validated in the MineRL environment within Minecraft, demonstrating a 12.6% reduction in Action Following Error on 300 human operation segments, with improvements reaching 20.9% on the most challenging segments. Odyssey, co-founded by Oliver Cameron and Jeff Hawke, recently secured investments from NVentures and Samsung Next, alongside existing backers GV, EQT, and Air Street Capital.