Google is expanding enterprise access to AI that helps robots plan tasks and adapt when conditions change. The challenge is turning that reasoning into reliable physical work.
Google Cloud announced on Oct. 27 that Gemini Robotics Embodied Reasoning (ER) 2 is available in public preview through Model Garden on its Gemini platform. The rollout expands enterprise access to robotics reasoning, with potential applications in warehouses, logistics, and laboratories.
The company introduced Gemini Robotics ER 2 on July 30, making it publicly available through the Gemini API and Google AI Studio, and offering it in private preview on Gemini Enterprise Agent Platform. The October announcement expands that enterprise offering to public preview.
“Advanced reasoning models have the potential to help robots interpret complex environments, plan actions and adapt to changing conditions, opening new possibilities for physical AI. Google’s Gemini Robotics ER 2 model represents an important step in this direction,” said Barrie Mullins, Senior Director of Arm Total Design for Physical AI, at Arm.
Google lists promotional pricing through Dec. 31, 2026, at US$1 per million input tokens, US$0.10 per million cached input tokens, and US$5 per million text output tokens, including reasoning. Google says the promotion is provided through 50% credits back on net spend. Starting Jan. 1, 2027, the listed rates double to US$2, US$0.20, and US$10, respectively.
How Gemini Robotics ER 2 works
Gemini Robotics ER 2 is a vision-language model that interprets visual information and uses natural language processing to understand instructions.
It uses video feeds to track task progress and identify critical moments. ER 2 handles high-level reasoning and planning, while lower-level vision-language-action models or robotics APIs handle execution.
One of the key features that this AI model brings is multi-robot collaboration in a shared physical space. According to Google, robots using the model can coordinate and hand off tasks to complete a shared workflow.
For enterprise buyers concerned about compatibility, a key differentiator that Gemini Robotics ER 2 has is that the same reasoning model can support different robot bodies. This means organizations could potentially reuse reasoning capabilities across different robots rather than develop entirely separate AI systems for each platform.
However, compatibility with different bodies does not automatically mean every robot can perform every task. The physical systems like sensors, actuators, and controllers still determine the robot's range of capabilities.
What eWeek found: Safety depends on more than robot reasoning
eWeek’s review of Google’s announcement and safety technical report highlights an important distinction: ER 2’s reasoning capabilities are only one part of a robot’s safety. The model can help plan tasks and recognize failures, but action models, control software, and physical hardware still determine how those decisions are executed.
Google reports 97.9% accuracy on its Safety Instruction Following benchmark and 93.0% on its Human Proximity benchmark at a one-meter threshold. These scores measure performance on defined evaluation tasks; they are not overall workplace-safety rates.
The company’s safety technical report explicitly says physical protections, including emergency stops, barriers, and speed or force limits, remain essential. The report also evaluates whether a reasoning model can reject unsafe requests, recognize tasks beyond an action model’s capabilities, and ask humans to clarify uncertain instructions.
For enterprises, that makes the complete system the unit that needs testing. A warehouse deployment would need to account for people entering a robot’s path and objects shifting unexpectedly. A laboratory deployment would need to account for unclear instrument readings, fragile equipment, and actions that cannot safely be repeated after a mistake.
A useful pilot should therefore establish when the robot stops, when it requests human help, and how operators intervene after an error. Recognizing a failed step is valuable, but retrying is appropriate only when the next attempt can be made safely — a challenge also illustrated by humanoid robots attempting a simulated firefighting mission.
The broader significance of the cloud preview is easier access to a reasoning layer that can support different machines. Whether that translates into useful automation will depend on how well each deployment combines reasoning, physical execution, safeguards, and human oversight.
Read more: Google’s cloud rollout builds on its broader robotics push—explore how Gemini Robotics 2 coordinates humanoid movement, task planning, and collaboration.