V-JEPA 2 and V-JEPA 2-AC

AI model
Humanoids and robot learning

Meta FAIR's PyTorch code for V-JEPA 2 video encoders and the V-JEPA 2-AC action-conditioned world model for robot manipulation, plus the V-JEPA 2.1 pretraining recipe.

Made by Meta, United States (checked on github.com/facebookresearch (opens in a new tab)).

Source on GitHub: github.com/facebookresearch/vjepa2 (opens in a new tab). Facts read at commit 204698b (opens in a new tab).

Licenses

Code: MIT (opens in a new tab) AND Apache-2.0 (opens in a new tab)
Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit 204698b (opens in a new tab), on 16 Oct 2026. A person at Robotheca set it on 5 Nov 2026.
Weights: Apache-2.0 (opens in a new tab)
Read on huggingface.co/facebook/vjepa2-vitg-fpc64-256 (opens in a new tab).
Commercial use: Yes

What it does: Robotheca's summary

Pretrains video encoders with a masked latent feature prediction objective on internet-scale video, without labels.
Provides pretrained ViT backbones from 300M to 2B parameters, loadable via PyTorch Hub or HuggingFace.
V-JEPA 2-AC is post-trained from the ViT-g encoder on a small amount of robot trajectory data.
V-JEPA 2-AC plans robot actions from image goals, reaching, grasping and pick-and-place with a Franka arm.
Includes attentive-probe evaluation on Something-Something v2, Diving48 and EPIC-KITCHENS-100, with training and inference configs.
V-JEPA 2.1 adds dense predictive loss, deep self-supervision and multi-modal tokenizers for dense features.

Written by Robotheca from the project's README (opens in a new tab) at commit 204698b.

Getting started

Difficulty: Advanced
Hardware it needs, as the project states it: Requires PyTorch, timm and einops; installing PyTorch with CUDA support is strongly recommended. The demo script assumes a GPU. macOS users need an alternative decord implementation since decord does not support macOS.
How to start: The project's README (opens in a new tab)

Maintenance

Status: Maintained: a commit in the last 12 months
Newest commit: 23 Mar 2026 on main
Latest release: None published on GitHub

Checked on GitHub on 31 Oct 2026.

Robots that can run it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.