World Model as a Graph: Learning Latent Landmarks for Planning
This work is significant for researchers and practitioners in reinforcement learning, as it offers a more scalable approach to long-horizon planning in complex environments, potentially enabling more robust and generalizable agents.
The paper addresses the challenge of long-horizon planning in model-based reinforcement learning, where traditional world models suffer from divergence over extended virtual rollouts. They propose a graph-structured world model, L3P, which learns latent landmarks as nodes and reachability estimates as edges, demonstrating significant performance improvements over prior methods on high-dimensional continuous control tasks.
Planning - the ability to analyze the structure of a problem in the large and decompose it into interrelated subproblems - is a hallmark of human intelligence. While deep reinforcement learning (RL) has shown great promise for solving relatively straightforward control tasks, it remains an open problem how to best incorporate planning into existing deep RL paradigms to handle increasingly complex environments. One prominent framework, Model-Based RL, learns a world model and plans using step-by-step virtual rollouts. This type of world model quickly diverges from reality when the planning horizon increases, thus struggling at long-horizon planning. How can we learn world models that endow agents with the ability to do temporally extended reasoning? In this work, we propose to learn graph-structured world models composed of sparse, multi-step transitions. We devise a novel algorithm to learn latent landmarks that are scattered (in terms of reachability) across the goal space as the nodes on the graph. In this same graph, the edges are the reachability estimates distilled from Q-functions. On a variety of high-dimensional continuous control tasks ranging from robotic manipulation to navigation, we demonstrate that our method, named L3P, significantly outperforms prior work, and is oftentimes the only method capable of leveraging both the robustness of model-free RL and generalization of graph-search algorithms. We believe our work is an important step towards scalable planning in reinforcement learning.