Computational Model Studio · Model release
OpenVLA 7B
Predicts robot manipulation actions from a camera image and an English-language instruction.
Developer: OpenVLA collaboration · Last verified:
What it is designed to do
Predict normalized 7-DoF robot end-effector actions from a workspace image and language instruction.
- Primary contribution
- control action
- Input
- image, text
- Output
- number vector (normalized 7-DoF robot action)
Documented uses: Action prediction for robot embodiments and settings represented in the Open-X training mixture.
Access and deployment
- OpenVLA 7B publisher weights — available. Public publisher repository; confirm MIT terms and runtime dependencies before use. Artifact documentation Local hardware requirements have not been verified. The full artifact-specific rights document has not been separately reviewed; follow the stated publisher or upstream license conditions.
Technical profile and review notes
- Construction
- learned
- Representation
- neural
- Normalized architecture tags
- No normalized architecture tag assigned
- Lifecycle
- Unknown from reviewed sources
Inclusion rationale: Fills the control/action gap with a bounded model output contract, distinct from a general robot control stack.
Limitations
- Predicted actions require per-robot, per-dataset un-normalization before execution.
- The publisher does not claim zero-shot generalization to unseen robot embodiments or setups.
- The model predicts an action; safe physical execution depends on the surrounding robot system.
Sources and related Studio records
Sources
Connections
- In areaComputational Models
- Supports functionPredict robot actions
- ArtifactOpenVLA 7B publisher weights