What is Gemini Robotics? Action models and embodied reasoning
Gemini Robotics is Google DeepMind’s robotics family, separating vision-language-action models from ER embodied reasoning. Learn about Robotics 2 inputs and outputs and why a preview API is not a complete robot controller.
Meet this model
Gemini Robotics is Google DeepMind’s family for physical robots. Vision-language-action models turn visual and language inputs into control, while ER embodied reasoning handles spatial understanding and planning. The family now includes Robotics 2 and ER 2, but calling a reasoning endpoint does not supply a complete controller for any robot.
Action generation and planning serve different roles
Putting a block into a box requires both localization and grasp execution. A VLA connects perception and instructions to actions; an ER model can interpret the scene, plan steps and coordinate tools. Integrating these results still requires the robot’s own control interface.
The Robotics 2 family also distinguishes an On-Device version. Deployment location and control frequency depend on the system design. A completed action in a research demonstration does not establish support for every robot, whose structure, sensors and motion range can differ.
Check coordinates and completion against the actual scene
ER 2 publicly supports pointing, spatial reasoning and video progress assessment. For a returned point or box, verify image dimensions and coordinate conventions before converting it for control. An image location is not automatically a reachable point in robot coordinates.
Confirming that the block entered the box requires subsequent sensor or visual evidence. Success labels can be wrong under occlusion or similar-looking objects. Start by checking recorded scenes; a detailed answer is not evidence that a physical action happened.
The ER preview is not a universal robot-control API
The current Gemini Developer API offers ER 2 previews with standard and streaming variants. Audiovisual inputs lead to text or structured descriptions for software. It can use developer-connected robot functions, but does not automatically discover and control hardware from arbitrary manufacturers.
ER 1.6 is marked deprecated, so new integrations should follow the current catalog and migration guidance. Public embodied-reasoning access does not establish unconditional general testing of the full VLA family. Check account and regional conditions; this article did not request or call these endpoints.
Physical actions require a controlled environment and independent stopping
Physical errors affect surrounding objects, so initial validation should limit motion and retain stopping controls outside the model. Correct spatial understanding can still encounter gripper or sensor failures, handled by the runtime. A chat-app model selector cannot replace robot integration.
Official documentation notes possible model errors and detection effects from lighting or low contrast. Check image sources and hardware adaptation for a real site. This article reviewed the model page and ER documentation without connecting a robot or grasping objects and makes no hardware-compatibility or success-rate claim.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.