The authors propose a method to learn temporal relations from random exploration trajectories, enabling goal-directed planning without requiring reward labels or policy improvement during training. A conditional energy-based model, trained via noise-contrastive estimation on observation pairs at multiple horizons, captures how likely a state is to lead to a goal within a given number of steps. At test time, a separate one-step dynamics model predicts action outcomes, and the temporal model scores their progress toward the goal, allowing the agent to replan after each action.
Experiments show that this approach supports long-range maze planning from random walk data using both state vectors and images, and extends to egocentric navigation and manipulation from suboptimal trajectories. The learned representations exhibit properties of a multiscale cognitive map, with score fields and embedding probes reflecting geodesic and connectivity structure. This work is interesting because it separates long-range guidance from local dynamics, enabling zero-shot goal reaching from exploration data without explicit reward engineering or policy training.