Google Cloud details Gemini Robotics ER 2

Google describes Gemini Robotics ER 2 as a specialized model that acts as a robot’s high-level brain.[1] The company says it combines spatial and temporal video understanding with long-horizon planning and built-in safety guardrails.[1] It says the model can handle tasks lasting multiple minutes and involving hundreds of decisions.[1] That description positions ER 2 as a reasoning layer for extended tasks, while developers can declare lower-level interfaces—including VLA models, navigation APIs or custom functions—as tools for it.[1]

Google says developers can stream video, audio or text directly into ER 2.[1] The model can classify task progress as a binary status, discrete categories or continuous scores, giving teams different ways to represent whether work is proceeding as intended.[1] On locating critical frames in video, Google reports 91.3% accuracy and a 0.96-second mean absolute frame distance, while saying the feature runs at four times the execution speed.[1] These are Google’s reported results; the evidence provided does not specify the evaluation setup.

Google says ER 2 can work across stationary arms, wheeled rovers and quadrupeds.[1] It says robots can use the model to share an understanding of their workspace, divide tasks and hand work to one another.[1] Teams can test robot behaviors in MuJoCo-Warp and NVIDIA Isaac Sim, and simulate, orchestrate and manage physical agents before deploying them on robots at the edge.[1] Google also describes using ER 2 features to curate and label demonstration videos in Cloud Storage and BigQuery.[1]

Google Cloud associates the offering with enterprise-grade security, multi-agent orchestration and end-to-end lifecycle management.[1] Google says operators can redirect the model without technical language, and that Google Distributed Cloud can offload latency-sensitive local workloads such as SLAM and path planning on factory floors.[1] Google DeepMind’s model card says ER 2 is based on Gemini 3.5 Flash, and advises against using its robotics models for safety-critical applications such as healthcare or transportation.[2] DeepReach, Arm and ASTAR IMRE representatives also described the model positively.[1]