WA-JEPA: Rethinking the Video JEPA Paradigm for World-Action Modeling in Autonomous Driving
Researchers have developed WA-JEPA, a new approach to autonomous driving planning that improves upon the existing Video Joint Embedding Predictive Architecture (V-JEPA). Instead of using random masking and deterministic regression, WA-JEPA employs hybrid future-masked pre-training and conditional flow matching over latent futures. This allows for more accurate generation of plausible future latents, which is crucial for downstream planning. The model was pre-trained on nuPlan
Researchers have developed WA-JEPA, a new approach to autonomous driving planning that improves upon the existing Video Joint Embedding Predictive Architecture (V-JEPA). Instead of using random masking and deterministic regression, WA-JEPA employs hybrid future-masked pre-training and conditional flow matching over latent futures. This allows for more accurate generation of plausible future latents, which is crucial for downstream planning. The model was pre-trained on nuPlan videos and fine-tuned on NAVSIM, achieving state-of-the-art results on several benchmarks. Code is available on GitHub.
---
Why it matters: This matters to engineers working on autonomous driving systems because WA-JEPA offers a more accurate and scalable approach to world-action modeling, which is essential for planning and decision-making in complex scenarios.
Source: https://arxiv.org/abs/2608.20974
This article was originally published at: https://arxiv.org/abs/2608.20974