V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

arXiv 25.6 2025 55.5 method

TLDR

Self-supervised video model V-JEPA 2 achieves strong video understanding and action anticipation, and enables zero-shot robotic planning via a latent action-conditioned world model.

Reasoning

Strengths: Combines internet-scale video with minimal interaction data, achieves SOTA on multiple tasks, and demonstrates real-world robot planning zero-shot. Weaknesses: Limited detail on the world model's generalization and scalability; reliance on LLM alignment for QA tasks may introduce complexity.

Read-first score

Read-first score 55.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 40.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,result

Topical relevance 42%

57.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 301.

Keyword Scores

world model

video world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

Deep Analysis

Innovations

Self-supervised approach combining internet-scale video data with a small amount of interaction data (robot trajectories) for understanding, prediction, and planning
Action-free joint-embedding-predictive architecture (V-JEPA 2) pre-trained on over 1 million hours of internet video and images
Alignment of V-JEPA 2 with a large language model to achieve state-of-the-art performance on multiple video question-answering tasks at the 8 billion parameter scale
Post-training a latent action-conditioned world model (V-JEPA 2-AC) using less than 62 hours of unlabeled robot videos for zero-shot robotic planning on Franka arms

Methodology

The paper pre-trains an action-free joint-embedding-predictive architecture (V-JEPA 2) on a large dataset of internet video and images (over 1 million hours). For video question-answering, V-JEPA 2 is aligned with a large language model. For robotic planning, a latent action-conditioned world model (V-JEPA 2-AC) is post-trained using less than 62 hours of unlabeled robot videos from the Droid dataset, then deployed zero-shot on Franka arms in two different labs without task-specific training or reward.

Key Results

V-JEPA 2 achieves 77.3 top-1 accuracy on Something-Something v2 for motion understanding and 39.7 recall-at-5 on Epic-Kitchens-100 for human action anticipation, surpassing previous task-specific models. After LLM alignment, it achieves state-of-the-art results on video QA (84.0 on PerceptionTest, 76.9 on TempCompass) at the 8B parameter scale, and the V-JEPA 2-AC model enables zero-shot picking and placing of objects using planning with image goals on Franka arms.

Resources

Paper Project