VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model
TLDR
VLA-JEPA uses a JEPA-style latent world model to predict action-relevant state transitions, improving VLA policy robustness and generalization.
Reasoning
The paper clearly identifies a key limitation of existing latent-action objectives (pixel variation bias) and proposes a leakage-free latent prediction method. Strengths include a simple two-stage recipe and demonstrated gains on multiple benchmarks including real-world tasks. Weaknesses are that the abstract does not detail limitations or ablation studies, and the novelty relative to prior JEPA work is not fully elaborated.
Read-first score
Read-first score 52.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.
Field roles
Rank sensitivity
Stability: volatile; rank range: 456.
Keyword Scores
Deep Analysis
Innovations
- JEPA-style pretraining framework for Vision-Language-Action models
- Leakage-free state prediction where future information is used only as supervision targets, never as input
- Predicting in latent space rather than pixel space to learn dynamics abstractions robust to camera motion and irrelevant background changes
- Simple two-stage recipe: JEPA pretraining followed by action-head fine-tuning, avoiding multi-stage complexity
Methodology
VLA-JEPA uses a target encoder to produce latent representations from future frames, while the student pathway sees only the current observation. Future information is used solely as supervision targets, never as input. The model is pretrained on video data with a JEPA-style objective in latent space, then fine-tuned with an action head for downstream tasks.
Key Results
Experiments on LIBERO, LIBERO-Plus, SimplerEnv, and real-world manipulation tasks show that VLA-JEPA achieves consistent gains in generalization and robustness over existing methods.