Title: Laplacian Representations for Decision-Time Planning

Abstract:

Developing robust planning strategies using learned models continues to present a significant hurdle in model-based reinforcement learning (RL). Within the context of decision-time planning, the quality of state representations is paramount; they must simultaneously facilitate local cost calculations and maintain the structural integrity of long-horizon trajectories. This study demonstrates that Laplacian representations serve as a potent latent space for planning, effectively encoding state-space distances across various temporal scales. By maintaining meaningful distance metrics and inherently breaking down long-horizon challenges into manageable subgoals, this approach helps alleviate the accumulation of errors typically associated with extended prediction windows. Leveraging these advantages, we propose ALPS, a hierarchical planning framework. Our experiments reveal that ALPS surpasses standard baseline methods across several offline goal-conditioned RL tasks within OGBench, a benchmarking suite that has historically been dominated by model-free approaches.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC