Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 17·5 min read
Traj-LeWM: Path-Aware World-Model Planning via Latent Trajectory Cost
TL;DR
Traj-LeWM enhances the existing LeWM model by incorporating a goal-conditioned latent trajectory cost (LTC) to improve planning accuracy. This approach leads to significant performance gains across various tasks, outperforming LeWM by up to 14 percentage points.
✦ Why It Matters
Consider integrating trajectory-level information into your planning models to improve decision-making accuracy.
Key Takeaways
- Incorporate goal-conditioned trajectory costs in planning models.
- Use trajectory-preference supervision during model training.
- Evaluate complete trajectories, not just endpoints, for better predictions.
- Test improvements on benchmarks like Push-T and Reacher.
How It Works
Traj-LeWM retains the local-dynamics learning of LeWM while adding a latent trajectory cost (LTC) that evaluates entire action sequences. This allows the model to consider not just where an action ends, but how it progresses towards the goal, providing a richer context for decision-making.