Traj-LeWM Improves World Model Planning with Path-Aware Cost

LeWM, a lightweight visual world model, learns latent dynamics directly from pixel data and evaluates potential action sequences by measuring the distance between their predicted endpoints and the desired goal. However, this approach has two primary limitations. During training, it focuses on local, next-step transitions without explicitly evaluating complete trajectories in relation to the overall task goal. Furthermore, in the planning phase, it relies solely on endpoint distance for ranking candidate actions. The issue with relying only on endpoint distance is that model predictions might not perfectly align with actual execution outcomes. A candidate action sequence whose predicted endpoint is closest to the goal might not be the most effective when executed in the real environment. The evolution of the entire predicted trajectory, therefore, offers valuable complementary information beyond just the final endpoint. To address these shortcomings, Traj-LeWM is proposed. It retains LeWM's local dynamics objective and endpoint scoring but integrates a novel goal-conditioned latent trajectory cost (LTC). This LTC aggregates trajectory-level information, serving as an additional signal. During training, LTC-based trajectory-preference supervision augments next-step prediction, shaping a more informed shared representation. In planning, LTC is combined with endpoint distance, allowing intermediate path information to influence candidate ranking. This joint scoring mechanism enables Traj-LeWM to significantly outperform LeWM across various benchmarks, demonstrating the benefits of path-aware planning.