Google Cloud opens Gemini Robotics ER 2 to enterprises in public preview
Google Cloud said in a blog post on Oct 27 that Gemini Robotics ER 2, Google DeepMind's embodied-reasoning model, is now in public preview on its Gemini platform through Model Garden. Google says the model finds the key frame of a physical event with 91.3% accuracy and a 0.96-second mean error, reads 10+ kinds of gauges and displays, and orchestrates VLA models and robot APIs as tools; Boston Dynamics, Arm, DeepReach and A*STAR IMRE are named as early users.
Google said in a Google Cloud blog post on Oct 27 (US time) that Gemini Robotics ER 2, the high-level reasoning model Google DeepMind released through the Gemini API on Jul 30, is now in public preview for enterprise customers on the Gemini platform, available through Model Garden.
The model acts as the planning layer above a robot's motor control. Developers declare low-level controllers such as vision-language-action (VLA) models or navigation APIs as tools, stream video, audio or text into ER 2, and let it break a task into steps, check each step and retry a failed one without restarting the workflow. Google gives two measured figures: ER 2 pinpoints the video frame of a critical event, such as when to stop pouring, with 91.3% accuracy and a 0.96-second mean absolute frame distance, and does so at four times the speed of many larger model categories. It also reads values from more than ten instrument types, from circular dials and sight glasses to rulers and liquid thermometers.
The cloud package links simulation in MuJoCo-Warp and NVIDIA Isaac Sim, data labelling in Cloud Storage and BigQuery, custom VLA training on GKE or Cluster Director, and edge deployment of latency-sensitive SLAM and path planning through Google Distributed Cloud. Quotes in the post come from Boston Dynamics (video inspection with Spot), Arm, the data company DeepReach AI and Singapore's A*STAR IMRE, which wants robots to read the touchscreen of a lab reactor that has no API.
MEKAWIRE context: this is the third Google robotics model release we have covered in four months, after ER 2 itself in July and Gemini 4 Argon for Robotics on Oct 13. The step now is commercial, not scientific: Google is selling the reasoning layer as a cloud service while hardware makers keep the motors, the split described in our VLA model explainer. A public preview carries no service-level commitment, and Google gave no pricing in the post.