Gemini Robotics ER 2 Reaches Enterprise Cloud in Public Preview Without Live API Streaming

Google made its Gemini Robotics Embodied Reasoning 2 model available in public preview on the Gemini Enterprise Agent Platform on October 27, giving enterprise teams their first direct cloud path to a physical AI reasoning layer without requiring a hardware partnership with Google. The model, accessible via Google Cloud's Model Garden under the identifier gemini-robotics-er-2-preview, handles high-level robot task planning and multi-robot coordination — but it arrives with a constraint that distinguishes it from the version Google released via the Gemini API three months earlier: real-time bidirectional streaming is absent from the cloud endpoint.

The October Cloud launch is not the same event as the July 30 Gemini API launch. Enterprise teams now accessing ER 2 through the Gemini Enterprise Agent Platform get a standard REST-based endpoint; the real-time streaming that appeared in the original Gemini API version is not available here. The distinction defines what the cloud version can and cannot do in practice.

ER 2 Handles Planning, Not Motor Commands

Understanding what ER 2 actually does requires distinguishing between two categories of model the robotics AI field often conflates. A vision-language model (VLM) processes images, video, and natural language and produces structured text output — descriptions, plans, decisions. A vision-language-action model (VLA) translates perception directly into motor commands that drive actuators. ER 2 is a VLM. It observes a scene, understands a task, and generates a multi-step plan; it does not produce the servo angles or joint torques that move a robot arm.

Google built ER 2 on Gemini 3.5 Flash, extending its vision capabilities with robot camera training data that allows the model to reason about task progression across time. According to Google DeepMind's model card, the model accepts up to 131,072 tokens of text and image input, or 65,536 tokens of video, and outputs structured task descriptions that a separate lower-level controller executes.

That lower-level layer — Google's Gemini Robotics 2, a full VLA capable of generating motor commands — remains accessible only to hardware partners. What the public preview opens is the planning brain; the action executor stays closed for now. Teams integrating ER 2 must pair it with their own robot control stack or a partner-provided actuation layer.

Live API Streaming Remains Unavailable in the Enterprise Cloud Version

The original Gemini API launch included a streaming endpoint (gemini-robotics-er-2-streaming-preview) that connects to Google's Live API — a persistent low-latency connection enabling bidirectional model interaction. That architecture allows a robot system to continuously feed camera data into the model and receive near-real-time instructions, which is essential for reactive behaviors: dodging an unexpected obstacle, adjusting grip when an object shifts, responding to human presence in a shared workspace.

The Enterprise Agent Platform version does not support the Live API. It operates on standard REST request-response patterns, a gap confirmed by XenoSpectrum's analysis of the two access paths. For use cases that tolerate that latency profile — task pre-planning before a workflow begins, simulation testing in MuJoCo-Warp or NVIDIA Isaac Sim, batch progress evaluation after tasks complete — the cloud endpoint is functional. For reactive manipulation and real-time navigation, developers still need the Gemini API endpoint from July. Teams evaluating ER 2 for production deployment must match their latency requirements to the correct access path.

A Three-Layer Physical AI Stack, Two Layers Still Partner-Only

Google's physical AI offering has three components targeting different points in the robot control hierarchy. Google Cloud's official announcement describes ER 2 as the top tier: scene understanding, long-horizon task sequencing, and multi-robot coordination. It is hardware-agnostic and designed to work with any platform that can accept structured task instructions.

The second tier, Gemini Robotics 2, is a full VLA that translates ER 2's plans into motor commands. Development partners include Apptronik (Apollo 2 humanoid), Boston Dynamics (Spot quadruped), and Franka Robotics (F3 Duo robotic arm), alongside Singapore's A*STAR Institute of Materials Research and Engineering. This model remains in partner-only access. The third tier, Gemini Robotics On-Device 2, targets edge inference for offline environments and is also partner-only.

The practical implication is that ER 2 alone cannot drive a robot. It produces a plan; something else must execute it — a constraint that defines the integration work required of any team building on the public preview.

Google's Benchmarks Show Promise, Independent Verification Still Absent

According to Google DeepMind's model card, on moment-finding — identifying when a specific step occurs within a task video — ER 2 achieved 91.3% accuracy with a mean absolute frame distance of 0.96 seconds. On progress classification, which assigns a task to one of five completion stages, it reached 57.4%. Both figures come from Google's own evaluation runs, not independent reproduction.

The moment-finding result is strong; the 57.4% progress classification is more modest and suggests meaningful room for improvement before ER 2 could serve as a reliable quality-control layer in production manufacturing. Google has not publicly disclosed results on ASIMOV2, a safety benchmark for robotics models. None of these metrics fully capture the failure modes that emerge when ER 2 operates through a VLA to a physical robot: novel lighting conditions, unexpected object positions, and recovery from failed manipulation attempts are harder to evaluate in controlled benchmarks.

How ER 2 Compares With NVIDIA's GR00T and the Rest of Physical AI

NVIDIA's GR00T N1.5, released at Computex 2025 in May, remains the closest publicly accessible comparable. GR00T N1.5 is a complete cross-embodiment VLA capable of edge inference on NVIDIA Jetson and available as open weights through NVIDIA's Isaac developer platform. It closes the perception-to-action loop in a single model rather than requiring a separate planning layer.

ER 2 targets a different scenario. Its value proposition is enterprise cloud integration: teams already on GCP gain access to embodied reasoning through the same IAM controls, VPC Service Controls, and audit logging that govern their other cloud workloads. For regulated environments or logistics operations with existing Google Cloud contracts, that context matters more than edge inference capability. The two systems are complementary rather than directly substitutable.

What the field does not yet have is an accessible, hardware-agnostic, fully integrated VLA that combines GR00T's cross-embodiment breadth with ER 2's enterprise cloud controls. Google's three-tier architecture is designed to provide that, but the commercial trajectory of the complete stack depends on when Gemini Robotics 2 — the VLA layer — becomes available beyond the current hardware partner network. That access decision, not any future benchmark release, is the milestone that will determine how significantly the October public preview reshapes enterprise robotics development.