Traj-LeWM: Path-Aware World-Model Planning via Latent Trajectory Cost
A recent study published on arXiv (2608.14125) presents Traj-LeWM, an advancement of the LeWM visual world model that integrates a goal-conditioned latent trajectory cost (LTC) to enhance planning capabilities. LeWM is a streamlined visual world model that learns latent dynamics directly from pixel data and evaluates potential action sequences based on the distance between predicted endpoints and the target. However, it faces two challenges: it assesses only local next-step transitions without considering full trajectories concerning the task goal, and it ranks candidates based solely on predicted endpoint distance, which may not reflect actual performance. Traj-LeWM preserves LeWM's local-dynamics objective and endpoint scoring while introducing the LTC, which consolidates insights from the entire predicted trajectory, offering additional signals beyond mere endpoint distance. The research team behind this method aims to bolster the reliability of model-based planning in visual contexts.
Key facts
- arXiv paper 2608.14125 introduces Traj-LeWM
- Traj-LeWM enhances LeWM with a goal-conditioned latent trajectory cost (LTC)
- LeWM learns latent dynamics end-to-end from pixels
- LeWM ranks candidate action sequences by predicted endpoint distance
- LeWM has limitations in training and planning
- Traj-LeWM retains LeWM's local-dynamics objective and endpoint score
- LTC aggregates information from the complete predicted trajectory
- The paper is available on arXiv
Entities
Institutions
- arXiv