What Drives Success in Physical Planning with JEPA World Models?
4 months ago
- #Machine Learning
- #Robotics
- #Artificial Intelligence
- Investigates success factors in physical planning with Joint-Embedding Predictive World Models (JEPA-WMs).
- Compares planning in input space vs. learned representation space for efficiency.
- Proposes a comprehensive study of model architecture, training objectives, and planning algorithms.
- Tests models in simulated and real-world robotic tasks, outperforming baselines like DINO-WM and V-JEPA-2-AC.
- Provides code, data, and checkpoints for reproducibility.