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.
Odyssey Unveils PROWL Framework to Enhance World Model Training
Avertissement : Le contenu proposé sur Phemex News est à titre informatif uniquement. Nous ne garantissons pas la qualité, l'exactitude ou l'exhaustivité des informations provenant d'articles tiers. Ce contenu ne constitue pas un conseil financier ou d'investissement. Nous vous recommandons vivement d'effectuer vos propres recherches et de consulter un conseiller financier qualifié avant toute décision d'investissement.
