Hierarchical World Models with Implicit Dynamics
ID-WM reports a 135% performance gain over a single-level world-model planner on long-horizon tasks.
The system separates choosing reachable subgoals from finding the precise actions needed to reach them. A high-level model proposes feasible state transitions, while a low-level world model solves for the actions. The authors say this lets it assemble new long-horizon plans from short segments beyond the behaviors seen during training. Across three navigation and manipulation suites, they report a 38.5% improvement over state-of-the-art policy and diffusion-based planners.
ArXiv · AI/CL/LG's note
The system separates choosing reachable subgoals from finding the precise actions needed to reach them. A high-level model proposes feasible state transitions, while a low-level world model solves for the actions. The authors say this lets it assemble new long-horizon plans from short segments beyond the behaviors seen during training. Across three navigation and manipulation suites, they report a 38.5% improvement over state-of-the-art policy and diffusion-based planners.
ArXiv · AI/CL/LG's note
score 5