What Is a VLA Model?

Action is the defining output

A VLA is relevant to robot control because its output represents actions, rather than only a textual description of what a robot could do. RT-2 is a documented example of representing actions with tokens. The actual action space still matters: a model may emit commands that require interpretation by another controller. Inspect the implementation instead of assuming that a natural-language answer can directly operate hardware.

Inspect what the model sees and controls

Record the visual inputs, any robot-state input, instruction format and output units. Check how camera placement and action normalization are handled. OpenVLA documents data and adaptation workflows, which is useful evidence of the integration surface. A checkpoint download alone does not specify camera calibration, gripper conventions, control timing or the complete runtime on a new robot.

Keep planning and action distinct

Some systems separate embodied reasoning from lower-level action generation. Google DeepMind’s family distinguishes these roles, including a local-execution model. A buyer should ask which component chooses the task sequence and which produces robot commands. Then identify the interfaces, failure handling and human approval required between them. Do not transfer a reasoning demonstration into a measured action-success claim.

Treat adaptation as part of the experiment

Record the base checkpoint, fine-tuning method, training data and evaluation checkpoint. OpenVLA and OpenVLA-OFT are related but distinct implementations; their names should not be merged into one benchmark row. Changes to training, action generation or inference can alter the evaluated system. Preserve those choices so an apparent model improvement can be assessed and reproduced.

Require a bounded success claim

Ask which tasks, objects, scene conditions and robot body were tested, how failures were counted and whether trials involved human resets or assistance. A LIBERO result is evidence within its simulation protocol; it is not an unattended deployment guarantee. For commercial use, add a target-site evaluation and a clear recovery procedure before interpreting research success as operational readiness.

Additional disclosed fields

Open research reference

OpenVLA publishes model code, checkpoints and fine-tuning paths

Primary specReviewed 2026-09-28

Sources 1

openvla/openvla official repository

openvla · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Action-token research

RT-2 represents robot actions as tokens

Primary specReviewed 2026-09-28

Sources 1

RT-2 research project

Google DeepMind · Publication date not disclosed · Source accessed 2026-09-28

Role separation

Gemini Robotics distinguishes VLA action models and embodied reasoning models

Primary specReviewed 2026-09-28

Sources 1

Gemini Robotics model family

Google DeepMind · Publication date not disclosed · Source accessed 2026-09-28

Relationships & deployments

Related reading is an editorial crosslink. Sourced connections describe relationships reported in the cited material. A link to a versioned profile does not establish compatibility with that version unless the connection note explicitly identifies it.

Record history & verified changes

A research review records when we checked a source. It does not mark a product launch or a new deployment.

Initial reviewed record. No subsequent field change has been recorded.