According to Google's announcement, the artificial intelligence model Gemini Robotics Embodied Reasoning 2 (known as ER 2) is now available in public preview on the Gemini platform. The model was developed by Google DeepMind and serves as an intelligence layer and a high-level brain for various robotic systems, aimed at enabling them to think, act, and adapt in physical environments. According to the publication, earlier this year the company introduced Gemini Robotics 2, and now the embodied reasoning model is being offered for the first time for enterprise production environments on Google Cloud, alongside enterprise-grade security capabilities, multi-agent orchestration, and end-to-end lifecycle management for simulating, evaluating, and deploying physical artificial intelligence at scale. The model is designed to operate across diverse robotic embodiments, ranging from stationary arms to full humanoids. It combines spatial and temporal video understanding with long-horizon planning and built-in safety mechanisms, enabling robots to converse with humans, orchestrate hardware interfaces and low-level action models, execute multi-minute tasks involving hundreds of decisions, and collaborate in robot teams.
Physical Task Verification, Real-Time Understanding, and Self-Correction
In real-world physical operations, a robot needs to know whether a specific step was successfully completed before moving on to the next step. Google reports that the Gemini Robotics ER 2 model adapts to novel and unfamiliar situations in real time by reasoning over continuous video feeds.
The model classifies task progress into binary status, discrete categories, or continuous scores, aiming to improve real-time situational awareness and enable the adjustment or re-execution of failed steps without restarting the entire workflow. According to company data, the model is capable of pinpointing the exact video frame where a critical event occurs (such as the exact moment when pouring a liquid must stop) with 91.3% accuracy and a mean absolute frame distance of 0.96 seconds. This performance occurs at four times the execution speed and at a fraction of the compute cost compared to many categories of larger models.
In addition, the model advances spatial reasoning capabilities, including visual question answering and instrument reading. The model is able to read and interpret values from more than ten categories of measurement instruments directly from visual input, including circular dials, sight glasses, digital displays, linear scales, rulers, and liquid-based thermometers.
A Unified Intelligence Layer Across Varied Robot Embodiments
Google notes that Gemini Robotics ER 2 orchestrates robots of different shapes and sizes, extending its operation from stationary arms and wheeled rovers to four-legged robots and complex structures.
Beyond this, the model enables multi-robot collaboration. This capability allows robots of different types to share a common semantic understanding of their workspace, divide labor among themselves, and hand off tasks from one robot to another, thereby completing workflows that a single machine cannot perform on its own.
Agentic Tool Orchestration and Natural Language Interaction
The model makes it possible to transform a robot into a physical agent capable of reasoning about the next step while simultaneously orchestrating actual actions. Developers can define low-level control interfaces—such as Vision-Language-Action (VLA) models, navigation interfaces, or custom functions—as tools within the system. This configuration functions across simulation environments, real-world robot control, or remote tele-operation by humans. Developers stream multimodal video, audio, or text directly into the model.
According to Google, Gemini Robotics ER 2 is capable of orchestrating application programming interfaces for navigation and manipulator movement directly from natural language commands. This allows the robot to explain its approach while performing an action and enables human operators to redirect it without using technical language.
Agent Development, Simulation, and Management in Google Cloud
According to Google's announcement, development teams and enterprise organizations use the Gemini platform to simulate, orchestrate, and manage physical agents, which are subsequently deployed to operate directly on physical robots at the edge using Google Cloud infrastructure:
- Fast simulation: Testing robot behavior before touching physical hardware by running simulations in environments such as MuJoCo-Warp and NVIDIA Isaac Sim. Developers can then generate synthetic scenarios using multimodal models and automatically label workcell imagery on the Gemini platform.
- Smart training: Utilizing the model's spatial pointing, moment finding, and progress classification capabilities to curate and label demonstration videos stored in Cloud Storage and BigQuery. Subsequently, custom VLA models are trained using the Gemini platform, Google Kubernetes Engine (GKE), or Cluster Director.
- Operating work environments: Deploying model workflows with enterprise Identity and Access Management (IAM), security controls, and built-in function calling. Latency-sensitive local workloads, such as Simultaneous Localization and Mapping (SLAM) and path planning, are offloaded to factory floors using Google Distributed Cloud, and telemetry data from the robot fleet is streamed directly into BigQuery.
Integration of the Model by Industry Partners and Organizations
The announcement highlighted statements from industry figures integrating the technology:
Marco da Silva, Vice President and General Manager of Spot at Boston Dynamics, noted that the company is seeing monthly growth among customers using AIVI-Learning, and that Gemini Robotics ER 2 provides additional value. According to him, video changes the nature of inspection, as the Spot robot does not merely check whether something looks right at a single point in time; rather, applying artificial intelligence models to video opens up new possibilities.
Chris Liu, co-founder and Chief Technology Officer at DeepReach, stated that their hardware and AI platform captures how skilled specialists work around the world and turns it into a data flywheel for model training. According to him, the model provides accurate task understanding for every recording, helping to locate the most important data faster and at a much larger scale.
Barrie Mullins, Senior Director of Arm Total Design for Physical AI at Arm, noted that advanced reasoning models have the potential to help robots interpret complex environments, plan actions, and adapt to changing conditions. He added that the model represents an important step in this direction, and that Arm is exploring together with Google how these capabilities can shape the next generation of computing platforms for robotics.
Prof. Loh Xian Jun, Executive Director at the ASTAR Institute of Materials Research and Engineering (ASTAR IMRE) and Executive Director of MARVEL-AI, stated that their goal is to accelerate materials research and development workflows to enable scientists to test more ideas rapidly. According to him, most laboratory automation today relies on rigid scripts for single, isolated tasks. Using the model provides robots with the visual and physical flexibility to work with existing infrastructure—such as reading a touchscreen on an older microwave reactor that lacks an API, balancing and loading a centrifuge, or moving samples safely through a busy laboratory—bringing them closer to self-driving laboratories.
Access and Developer Resources
Developers and enterprise teams can begin working with Gemini Robotics ER 2 in preview through Model Garden on the Gemini platform. Google provides documentation and pricing, developer guides featuring sample code, and instructions for bidirectional streaming in the API documentation, alongside access to performance benchmarks and a technical safety report in Google DeepMind's publication.