SALT: Learning State- and Temporally-Abstracted World Models for Offline Long-Horizon Decision-Making

William Huang, Benjamin Freed

Keywords: temporal abstraction, state abstraction, offline reinforcement learning, model-based reinforcement learning, world models

Abstract: Long-horizon decision making requires agents to reason over temporally extended behaviors rather than individual low-level actions. Existing skill-based methods reduce the effective planning horizon but often plan in the raw environment state space, while latent world models learn compact representations but typically lack temporal abstraction. We introduce State-Abstracted Latent skills for Temporal planning (SALT), a framework that jointly learns temporal and state abstractions from offline trajectories to support planning in a compact latent space. In preliminary experiments on AntMaze-medium-diverse, SALT achieves $71.7 \pm 5.1$\% offline success after 50 epochs without model or planner fine-tuning. With online finetuning, SALT improves to $78.3 \pm 4.7$\% success and outperforms PPO and SAC trained from scratch under the same episode budget.

Submission Number: 147