Hierarchical Planning with Latent World Models
Hierarchical Planning with Latent World Models (HWM) enables robots to perform complex manipulation tasks from a single image goal without prior training for that specific task. It uses two world models operating at different time scales, where a high-level planner sets abstract goals and a low-level planner executes primitive actions. This hierarchy allows the system to handle long-horizon tasks effectively, achieving zero-shot success in real robots.