VISION-LANGUAGE-ACTION

A VLA turns multimodal context into robot action.

Vision-Language-Action models extend multimodal AI toward physical control: they consume visual and language context and produce actions or motor-control outputs for embodied systems.

CITATION RECORDCanonical source ↗
MATURITYAUTHORITY EXPLAINER + OPENRGD CANDIDATE BOUNDARY
LAST VERIFIED22 Sep 2026
SOURCE VERSIONCurrent Gemini Robotics / GR00T public materials + Agent Contracts 0.1.0
EVIDENCE BOUNDARYVLA is a model-family term; the OpenRGD ActionIntent boundary is candidate and non-normative.

WHAT VLA MEANS

Vision and language become inputs; action becomes an output modality.

Google DeepMind describes Gemini Robotics as a VLA family that turns visual information and instructions into robot actions, while NVIDIA describes GR00T N models as VLA models for humanoid robot learning and control. The exact action representation, temporal horizon and control surface differ across systems, so “VLA” names a model family rather than one universal interface.

MODEL LAYERS

Reasoning and action generation do not have to be the same model.

01

Perception

Images, video and other sensor-derived context describe the current environment.

02

Language

Instructions and task context specify goals, constraints or interaction.

03

Embodied reasoning

A higher-level model can understand physical space and plan multi-step work.

04

VLA policy

The action model converts multimodal context into robot actions or motor-control outputs.

05

On-device execution

Some VLA designs are optimized to run locally where latency or connectivity matter.

06

Adaptation

Modern systems increasingly target multiple robot embodiments rather than one fixed body.

MULTI-EMBODIMENT

Portability increases the importance of explicit body context.

Recent robotics models explicitly target different robot shapes and hardware. DeepMind’s current On-Device 2 materials, for example, describe adaptation to new robot embodiments and sensors. That makes a hidden assumption dangerous: the same high-level task does not imply the same reachable actions, sensing, payload or safety envelope on every body.

OPENRGD CANDIDATE BOUNDARY

A cognition provider may be a VLA without becoming the hardware authority.

In the current OpenRGD candidate Cognition Contract, a cognition provider may be an LLM, VLM, VLA, world model or deterministic planner. The proposed boundary is deliberately model-agnostic: a provider emits explicit action semantics and evidence references rather than gaining direct actuator authority by default.

01Vision + language
02VLA / cognition provider
03ActionIntent
04Somatic Translator
05CapabilityPlan
06Safety gate
07Body Adapter
08Hardware

IMPORTANT LIMIT

This is an OpenRGD candidate architecture, not a universal rule for all VLA systems.

Some production or research VLA systems output motor-control actions directly. OpenRGD’s candidate architecture introduces an explicit semantic and safety boundary because it is optimizing for inspectability, heterogeneous embodiments and downstream accountability. Those contracts remain candidate and non-normative until promoted through OpenRGD governance.

PRIMARY SOURCES