🏆 World-in-World: Building a Closed-Loop World Interface to Evaluate World Models
Agent's View
🧠
Instruction:
Navigate to the Toaster in the room and be as close as possible to it.
🦾
Environment Step 4-7:
Planning:
- Move leftward by 0.25.
- Move leftward by 0.25.
- Move forward by 0.25.
- Move forward by 0.25.
Method | #Param. | Input Type | Control Type | Model Type | Mean Traj. ↓ | Acc. ↑ |
|---|---|---|---|---|---|---|
PathDreamer [36] | 0.69B | RGB-D; Sem; Pano | Trajectory | Video Gen. Post-Train | 5.898 | 60.98 |
World-in-World: Building a Closed-Loop World Interface to Evaluate World Models
This leaderboard showcases performance metrics across different types of AI models in world modeling tasks:
Model Categories
- VLM: Vision-Language Models
- Image Gen.: Image Generation Models
- Video Gen.: Video Generation Models
- Video Gen. Post-Train: Post-training specialized Video Generation Models
Metrics Explained
- Acc. ↑: Accuracy score (higher values indicate better performance)
- Mean Traj. ↓: Mean trajectory error (lower values indicate better performance)
Notes
- † indicates post-training specialized models
- XXX indicates results pending/unavailable
- – indicates not applicable or not available
Results represent performance on world modeling evaluation benchmarks and may vary across different evaluation settings.