Yann LeCun's team proposed a hierarchical planning method to enhance JEPA's world model long-term reasoning ability

ME News Report, April 8 (UTC+8), recently, Yann LeCun’s team proposed a hierarchical planning method based on multi-time-scale latent world models, aiming to address two major challenges faced by learning-based world models in long-term control: the accumulation of prediction errors and the exponential growth of the search space. The method learns latent world models at different time scales and performs cross-scale hierarchical planning, enabling long-term reasoning while significantly reducing planning complexity during inference. This framework can serve as a plug-in abstraction module, suitable for various latent world model architectures and domains. Experiments show that in real-world non-greedy robot tasks (such as grasping and placing), given only the final goal, hierarchical planning achieved a 70% success rate, while single-layer world models had a 0% success rate. In physics-based simulation environments (such as pushing operations and maze navigation), hierarchical planning not only achieved higher success rates but also reduced the required planning computation time by up to three times. The method does not rely on task-specific rewards or externally provided sub-goals, demonstrating strong generalization capabilities in unseen environments and tasks. (Source: InFoQ)