Google Cloud has made Gemini Robotics ER 2 available today in public preview through Model Garden on the Gemini platform, giving enterprise developers access to a model for planning and coordinating robot work. Google says the model can use video to track physical tasks, direct robot controls and let people change instructions through ordinary conversation. Pricing was not disclosed.
The Cloud offering combines the reasoning model with enterprise security controls, orchestration for multiple agents and tools for managing deployments throughout their lifecycle. Developers can connect movement systems, including navigation APIs and Vision-Language-Action (VLA) models, as tools, then supply the model with text, audio or video. Google says this allows robots to describe their intended actions while carrying them out and accept an operator's verbal redirection.
Read more: Google Cloud Adds Three OpenAI Models in Private Preview
The model itself predates this Cloud preview: Google DeepMind launched Gemini Robotics ER 2 in July, with developer access through Google AI Studio and the Gemini API. That launch also included a private preview through the Gemini Enterprise Agent Platform and a demonstration in which Boston Dynamics Spot retrieved objects in response to natural language requests.
Google says Gemini Robotics ER 2 can assess whether a task has succeeded and retry an unsuccessful step without beginning the whole process again. For identifying the video frame containing a critical event, Google reports 91.3% accuracy and a mean absolute frame distance of 0.96 seconds, citing stopping a pour as one application. The company also says the model can interpret 10 or more categories of instruments, including digital displays and circular dials.
Google says the same reasoning layer supports machines ranging from fixed robot arms to humanoids, with different robots able to allocate work and pass tasks between them. Teams can test behaviors in NVIDIA Isaac Sim or MuJoCo-Warp before using physical hardware. Developers can also use the model to label recorded demonstrations stored in BigQuery and Cloud Storage, then train custom VLA models.
For operation, Google offers function calling and enterprise IAM controls through the Gemini platform. It says Google Distributed Cloud can handle local workloads sensitive to delay, such as path planning, while BigQuery receives telemetry from robot fleets.
