V-JEPA 2-AC

Model type

Note The developer does not label it vision-language-action, world action model or embodied reasoning: "V-JEPA 2-AC is a latent action-conditioned world model post-trained from V-JEPA 2 (using a small amount of robot trajectory interaction data) that solves robot manipulation tasks without environment-specific data collection or task-specific training or calibration."

Controls

Note The model predicts video representations; actions are found by planning against an image goal: "We deploy V-JEPA 2-AC zero-shot on Franka arms in two different labs and enable picking and placing of objects using planning with image goals."

Inputs

Paper: the predictor is trained on "predicting the representations of future video frames conditioned on past video frames, actions, and end-effector states"; the real-robot tests use a camera input, "from a monocular RGB camera zero-shot in a new environment". Video = camera, end-effector states = robot state.

Access

README, V-JEPA 2-AC: "Our action-conditioned checkpoint was trained from the ViT-g encoder." The README links this ViT-g/16 checkpoint for download.

Parameters

Note Meta gives one number per component, none for the whole: the action-conditioned predictor "Our action-conditioned world model is a 300M-parameter transformer network"; the encoder is one of "parameter counts ranging from 300 million (ViT-L) to 1 billion (ViT-g) parameters"; "Our action-conditioned checkpoint was trained from the ViT-g encoder."

Base model

"Our action-conditioned checkpoint was trained from the ViT-g encoder." README: "V-JEPA 2-AC is a latent action-conditioned world model post-trained from V-JEPA 2 (using a small amount of robot trajectory interaction data) that solves robot manipulation tasks without environment-specific data collection or task-specific training or calibration."