What happened
On October 27, 2026, Google Cloud announced public-preview availability of Gemini Robotics Embodied Reasoning 2 on its Gemini platform. Google describes ER 2 as a specialized Gemini model that acts as a robot’s high-level brain. It is intended to work across robot embodiments, from stationary arms to humanoids, combining spatial and temporal video understanding with long-horizon planning and built-in safety guardrails. The company says the model can support natural-language interaction, orchestrate low-level action models and hardware APIs, execute multi-minute tasks involving hundreds of decisions, and participate in multi-robot workflows.
The product framing is specifically agentic. A developer can declare low-level control interfaces as tools, including Vision-Language-Action models, navigation APIs, and custom functions. Video, audio, or text can be streamed to ER 2. The model then reasons about the next step while orchestrating those interfaces across simulation, real-world control, or human tele-operation. Google says an operator can issue a natural-language instruction, receive an explanation of the proposed approach while the action occurs, and redirect the system without technical commands.
Google emphasizes verification rather than merely action selection. ER 2 can classify task progress using binary states, discrete categories, or continuous scores, according to the announcement. It can also identify the video frame in which a critical event occurs, such as when liquid pouring should stop. Google reports 91.3% accuracy and a 0.96-second mean absolute frame distance for that task, while saying the system runs at four times execution speed and at a fraction of the compute cost of many larger model categories. The post also says it can interpret visual readings from more than ten categories of instruments, including circular dials, digital displays, rulers, sight glasses, and liquid thermometers.
The Cloud availability supplies the surrounding deployment workflow. Google positions the Gemini platform for simulating, orchestrating, and managing physical agents before they run on edge robots. Teams can test behaviors in MuJoCo-Warp or NVIDIA Isaac Sim, generate synthetic scenarios, and auto-label workcell imagery. The company says ER 2’s pointing, moment-finding, and progress-classification capabilities can help curate and label demonstration video in Cloud Storage and BigQuery, while custom VLA models can be trained through the Gemini platform, GKE, or Cluster Director. Workflows use enterprise IAM, security controls, and built-in function calling; latency-sensitive local work such as SLAM and path planning can be handled through Google Distributed Cloud, with fleet telemetry streamed into BigQuery. The announcement does not state a price, quota, region list, or general-availability date.
What the archive adds
This is a new access and deployment fact on a known story. Our August Telegram coverage described Gemini Robotics 2 as a family including ER 2 and an on-device model; this announcement makes the ER 2 component available in public preview through Google Cloud. The Field Note archive’s recent Gemini API CLI and Gemini 4 coverage shares the Google and Gemini names, but is not a direct predecessor to this robotics release.
Trend signal
This supports Agent-ready software exposes capabilities. ER 2 is not presented as a self-contained robot application: developers expose navigation, VLA, and custom control interfaces as callable tools, while the model coordinates them. Its progress classification and event detection also support Goal-scoped loops replace manual continuation, because the system is being given observable task-state signals instead of relying only on a model’s judgment that work is complete.
Why it matters
Physical-agent teams often face a split architecture: a model interprets an instruction, while separate systems handle perception, control, simulation, telemetry, permissions, and incident review. ER 2’s public preview does not remove that engineering work, but it offers a common reasoning and orchestration layer inside a cloud environment already equipped with IAM, function calling, data services, and simulation connections.
The more meaningful design choice is verification. A robot that can initiate a sequence is not necessarily one that can safely recover when a grasp, pour, inspection, or handoff fails. A model that consumes continuous video and produces task-progress states can give a workflow a concrete condition for retrying, escalating, or moving on. Teams should evaluate that loop against their own hardware and hazards rather than treating the reported frame-localization number as a general reliability guarantee.
What happens next
- Google will add more production-readiness detail, such as supported regions, pricing, quotas, or a general-availability path. (3 months, 65%). Watch for: preview documentation and product-pricing updates.
- Early users will test ER 2 most visibly in simulation-to-deployment workflows that pair it with existing control APIs rather than replace low-level control outright. (6 months, 70%). Watch for: customer architecture disclosures using declared tools and simulation environments.
Limits and uncertainty
The announcement is a vendor product post, and its performance figures are Google’s own claims. It does not describe the evaluation dataset, baseline systems, task distribution, failure modes, safety validation procedure, latency by deployment setting, or performance on specific robots. Public preview availability is also not a general-availability commitment. The source says the model has built-in safety guardrails and can work with Cloud security controls, but it does not establish that those controls are sufficient for any particular factory, laboratory, or public environment. Pricing, capacity, regional availability, and service limits are absent.
Skill to build now
Choose one physical workflow with a clear observable completion condition, such as “the valve reached its target position” or “the inspection image contains a readable value.” This week, map its control primitives as narrowly scoped tools, define the video or telemetry evidence that verifies each step, and specify what happens after a failed verification: retry, request operator review, or halt. Run that contract in simulation before attaching it to hardware. The useful artifact is not a broad robot prompt; it is a testable boundary between planning, tool execution, and evidence of completion.
Reader takeaway
ER 2’s public preview is a concrete enterprise-access milestone for Google’s robotics reasoning model, not proof that general-purpose robots are ready for unsupervised work. Its most consequential idea is architectural: connect high-level multimodal planning to permissioned control tools and continuous evidence that each physical step actually happened.
Open question
Will Google publish enough detail about ER 2’s task evaluations, latency, failure recovery, and preview economics for teams to compare it meaningfully with a narrower, in-house robotics stack?