Gemini Robotics ER 2 enters public preview on Google Cloud

A high-level brain for many kinds of robots

Gemini Robotics ER 2 is a specialized Gemini model and the core of Gemini Robotics 2, which Google introduced earlier this year as an intelligence layer for robots. Google says the model works across any robot embodiment, from stationary arms and wheeled rovers to quadrupeds and full humanoids.

According to the company, this lets robots converse naturally with humans, orchestrate low-level action models and hardware APIs, execute multi-minute tasks involving hundreds of decisions, and collaborate across multi-robot teams. In those teams, different types of robots can share an understanding of their workspace, divide labor and hand off tasks to complete workflows a single machine could not do alone.

Developers can declare low-level control interfaces, such as Vision-Language-Action (VLA) models, navigation APIs or custom functions, as tools. They can then stream video, audio or text into the model. Google says the model can orchestrate navigation and manipulator movement from natural language commands, explain its approach while acting, and let operators redirect it without technical language.

Verifying tasks and correcting mistakes

Google stresses that a robot must know whether a step succeeded before it starts the next one. Gemini Robotics ER 2 reasons over continuous video feeds to judge task progress, classifying it as a binary status, discrete categories or continuous scores. This allows adjustment or a retry of a failed step without restarting an entire workflow.

The company reports that the model can pinpoint the exact video frame where a critical event occurs, such as when to stop pouring a liquid, with 91.3% accuracy and a 0.96-second mean absolute frame distance. Google says it does so at 4x the execution speed and a fraction of the compute cost of many larger model categories.

The model also reads instruments. According to Google, it can interpret values across 10+ instrument categories, including circular dials, sight glasses, digital displays, linear scales, rulers and liquid thermometers, directly from visual input.

Simulation, training and deployment on Google Cloud

Google describes a workflow in three stages on the Gemini platform:

• Simulate: test robot behaviors before touching hardware in environments such as MuJoCo-Warp and NVIDIA Isaac Sim, then generate synthetic scenarios with multimodal models and auto-label workcell imagery.
• Train: use the model's spatial pointing, moment finding and progress classification to curate and label demonstration videos in Cloud Storage and BigQuery, then train custom VLA models through the Gemini platform, GKE or Cluster Director.
• Operate: deploy workflows with enterprise IAM, security controls and built-in function calling, offload latency-sensitive workloads such as SLAM and path planning to factory floors with Google Distributed Cloud, and stream fleet telemetry into BigQuery.

Google says this is the first time organizations can pair its latest embodied reasoning model with the security, multi-agent orchestration and lifecycle management of Google Cloud.

Early customer comments

Several organizations are quoted in the announcement. Marco da Silva, VP and General Manager of Spot at Boston Dynamics, said that "the ability to apply AI models to video opens a whole new world of possibilities" for inspection. Chris Liu, Co-founder and CTO at DeepReach, said the model gives the company accurate task understanding of every recording to help find the most relevant data.

Prof. Loh Xian Jun of ASTAR IMRE said the institute is using the model to give lab robots flexibility to work with existing infrastructure, such as reading a touchscreen on a legacy microwave reactor that has no API. Arm's Barrie Mullins described the model as "an important step" for physical AI.

What it means for developers

Developers and enterprise teams can start building today in preview via Model Garden in the Gemini platform. Google points to documentation, pricing, sample code and bidirectional streaming guides in the Gemini API Robotics documentation, as well as benchmarks and a safety technical report in the Google DeepMind announcement. As the model is in public preview, teams can begin testing it in simulation and on their own robots, with the surrounding cloud tooling covering the path from experiments to production.