VLA VS VLM

A VLM understands multimodal content. A VLA learns to produce robot actions.

The key difference between a Vision-Language Model and a Vision-Language-Action model is the learned output behavior: VLMs generate semantic multimodal outputs, while VLAs extend that representation into an embodiment-specific action space.

CITATION RECORD
MATURITYVLA VS VLM CONCEPT REFERENCE
LAST VERIFIED27 Sep 2026
SOURCE VERSIONOpenVLA + SmolVLA + Gemini Robotics primary materials
EVIDENCE BOUNDARYVLA and VLM are model-family terms; some systems reuse a VLM backbone inside a larger action policy.

SHORT ANSWER

The action modality is the dividing line.

Both VLMs and VLAs can process images and language. A VLM may caption a scene, answer a question, identify an object or reason about spatial relationships. A VLA is trained so that the multimodal representation also conditions robot actions.

COMPARISON

Same multimodal foundation, different operational role.

VLM

Typical inputs

Images or video plus text.

VLM

Typical outputs

Text, labels, coordinates, embeddings or other semantic representations.

VLM

Training emphasis

Large-scale image-text and multimodal understanding objectives.

VLA

Typical inputs

Images, language and often robot proprioception or embodiment state.

VLA

Typical outputs

Action tokens, continuous action vectors or temporal action chunks.

VLA

Training emphasis

Robot trajectories paired with observations and task instructions, often combined with prior VLM pretraining.

ARCHITECTURE

Many VLAs contain a VLM rather than replacing it.

OpenVLA starts from a pretrained vision-language architecture and adapts it to predict robot actions. SmolVLA explicitly combines a VLM backbone with an action expert. GR00T similarly uses a VLM to encode visual and language context while a separate action path processes robot state and actions.

WHY VLM ALONE IS NOT ENOUGH

Knowing where the cup is does not define how this robot should move.

A multimodal model can identify a cup and understand the instruction “put it in the sink,” but robot execution still depends on reachability, kinematics, gripper state, camera geometry, action representation and the target control interface. VLA training binds semantic understanding to examples of embodied action.

WHY VLA IS NOT THE WHOLE ROBOT

Adding action output still does not eliminate control, safety or hardware interfaces.

A VLA may predict low-level actions, but physical deployment still has responsibilities outside the learned model: sensor timing, calibration, actuator constraints, controller stability, emergency behavior and application-specific safety. VLA and VLM describe model roles, not complete robot architectures.

OPENRGD VIEW

VLM and VLA can both be cognition providers; the downstream contract determines authority.

In the OpenRGD candidate architecture, a VLM might produce semantic interpretation or a high-level proposal, while a VLA may produce a more action-oriented intent. Neither label alone determines whether the component has direct hardware authority.

PRIMARY SOURCES