VLA Gallery / All models / Industry models

V-JEPA 2-AC

Meta FAIR·Mila · 2025-06

VAIndustry modelsEncoder up to 1BOther

Trains a world model self-supervised on over 1 million hours of internet video, then attaches an action-conditioned predictor using under 62 h of unlabeled robot video.

Paper

V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and PlanningarXiv 2506.09985

Overview

Category

Industry models

Type

VA

Released

2025-06

Organization

Meta FAIR·Mila

Open weights

Yes

Parameters

Encoder up to 1B (paper)/1.2B (blog); AC predictor about 300M

Architecture

Backbone LLM

None

Vision encoder

V-JEPA 2 ViT-g (up to 1B)

Action head

Goal-image-based MPC planning (Other)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Self-supervised video encoder ViT (up to ViT-g) + action-conditioned predictor (about 300M, 24 layers, block-causal attention); goal-image-based MPC planning (no language input)

Training

Pretraining accelerators

Not disclosed (only a GPU-days comparison table on A100)

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed

Data type

Human video (internet video) + images + unlabeled real-robot video

Data scale

1M+ h of internet video + 1M images (VideoMix22M); AC post-training on under 62 h of unlabeled DROID robot video

Main datasets

VideoMix22M, DROID

Tasks & evaluation

Zero-shot pick-and-place success rate in new environments·new objects 65–80% (blog); deployed on Franka arms in two different labs

Robot

Franka single arm

Inference

Inference device & speed

MPC planning runs on a single RTX 4090; planning time not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

2.4 GB

Where it runs

Local PC

Reinforcement learning & base

Reinforcement learning

Not used · Self-supervised video pretraining + action-conditioned predictor training. Execution by MPC planning

Base control

Fixed · No base control (Franka arm, MPC planning)

Notes

Key features

Trains a world model self-supervised on over 1 million hours of internet video, then attaches an action-conditioned predictor using under 62 h of unlabeled robot video. Performed zero-shot pick-and-place of new objects at 65–80% on Franka arms in two labs.

Notes

Conflicting parameter statements: paper encoder 'up to 1 billion' vs blog '1.2 billion-parameter model'; parameters (B) use the larger value 1.2. Data hours (h) is the lower bound of internet video 'over 1 million hours'. Announced 2025-06-11.

Verification

Confirmed