Google Opens Gemini Robotics ER 2 in Public Preview on Google Cloud

Google Cloud has made Gemini Robotics ER 2, a model that acts as the planning layer for robots, available in public preview. The post of 28 October 2026, by Carolina Parada, vice president and head of DeepMind Robotics, and Adrian Jones of Google Cloud, says ER 2 was introduced as part of Gemini Robotics 2.[1]

Google DeepMind first introduced ER 2 on 30 July 2026. That post called it a high-level brain for robots that hands motor execution to a lower-level vision-language-action model, or VLA.[2][1]

Google describes ER 2 as a specialised Gemini model that works across robot types, from stationary arms to wheeled rovers, quadrupeds and humanoids. It checks whether a task is progressing, scores it and can retry a failed step without restarting the workflow. It can also find a key moment in a video, such as when to stop pouring, and read instruments including dials, digital displays and thermometers.[1]

Developers can declare VLA models, navigation APIs or custom functions as tools, and the model accepts streamed video, audio or text. Google says it can run multi-minute tasks involving hundreds of decisions, with built-in safety guardrails.[1]

The figures are Google’s own. The Google Cloud post gives 91.3% accuracy and a 0.96-second mean absolute frame distance for finding critical frames. It says the model runs at four times the execution speed and a fraction of the compute cost of many larger models. The 30 July post repeats the moment-finding result and adds 57.4% accuracy on progress classification, which it says beats previous and competing frontier models. Comparison values appear only in charts.[1][2]

On safety, Google calls ER 2 its safest model, citing gains on two benchmarks, Safety Instruction Following and Human Proximity. It says that in testing the model halted a humanoid robot when a person was nearby and resumed once the area was clear. We did not read Google’s technical safety report.[2]

The Google Cloud post ties ER 2 to a toolchain. It lists simulation with MuJoCo-Warp and NVIDIA Isaac Sim, demonstration video curated in Cloud Storage and BigQuery, and custom VLA training through the Gemini platform, GKE or Cluster Director. Latency-sensitive jobs such as SLAM and path planning run on Google Distributed Cloud at the edge.[1]

The status needs care. The 30 July post described public access through the Gemini API and Google AI Studio and a private preview on the Gemini Enterprise Agent Platform. The Google Cloud post says public preview and does not say how that relates to the earlier private preview. It names no regions or quotas, and a preview model is not a general availability release.[2][1]