On July 30, 2026, Google DeepMind released Gemini Robotics 2, a suite of three AI models that together form what the company calls “the intelligence layer for the next generation of truly adaptable robots.” The launch moves Google’s physical AI ambitions from proof-of-concept demos to a system enterprise teams can access today, most directly through the Gemini Enterprise Agent Platform.
For enterprise leaders tracking where AI is heading, this release carries a signal that goes beyond robotics: the same agentic reasoning infrastructure powering software agents is now being applied to physical systems, and the boundary between digital and physical automation is beginning to dissolve.
Three Models, One Stack
Google DeepMind structured the launch as three interlocking models, each solving a different layer of the physical AI problem.
| Model | Type | Primary Role | Availability |
|---|---|---|---|
| Gemini Robotics 2 | Vision-Language-Action (VLA) | Full humanoid and bi-arm control | Early-access hardware partners |
| Gemini Robotics ER 2 | Embodied Reasoning (VLM) | High-level planning, task orchestration, multi-robot coordination | Gemini API, Google AI Studio, Gemini Enterprise Agent Platform (private preview) |
| Gemini Robotics On-Device 2 | Efficient VLA | Local execution, rapid adaptation to new embodiments | Early-access hardware partners |
This architecture separates concerns the way enterprise software architects will recognize: ER 2 acts as the “brain” that reasons and orchestrates, while the VLA models handle motor execution. Enterprise teams adopting the ER 2 layer today can plug in updated control models as hardware evolves, without rebuilding the reasoning layer.
Whole-Body Control: From Torso to Toes
Previous versions of Gemini Robotics focused on upper-body manipulation, handling tabletop tasks where a robot’s hands and arms were the primary control surface. Gemini Robotics 2 extends this to full humanoid motion, including walking, crouching, stretching, and whole-body reach in cluttered spaces.
According to Google DeepMind’s official announcement, the model “can enable a humanoid to walk, crouch, stretch, and manipulate objects to clean up a cluttered room” and can control robots “from feet to fingertips.” The company demonstrated this across multiple robot embodiments using the same model checkpoint: the Apptronik Apollo 2 with two different hand configurations, and the Franka Duo with a Robotiq gripper.
The implication for enterprise deployment: a single trained model generalizes across different hardware, which reduces the per-deployment engineering cost that has historically made enterprise robotics projects prohibitively expensive.
Embodied Reasoning: The Agentic Layer for Physical AI
Gemini Robotics ER 2 is the component most directly relevant to enterprise teams today, because it is the piece available now through the Gemini API.
Think of it as the agentic orchestration layer for physical environments. The model accepts continuous video streams, audio, and text as inputs. Developers declare low-level control interfaces (VLA models, navigation APIs) as tools, which ER 2 calls in sequence to complete multi-step physical tasks. This mirrors the tool-use architecture that enterprise teams already know from software agents.
According to Google’s developer blog, ER 2 can natively call tools like Google Search and any user-defined function, enabling robots to look up real-world information mid-task. The model integrates with the Gemini Live API’s bidirectional streaming for low-latency orchestration, eliminating the “stop-and-think” pauses that made earlier embodied reasoning models impractical for time-sensitive operations.
A concrete example from the launch: Boston Dynamics Spot, orchestrated by Gemini Robotics ER 2, fetches objects on natural-language commands. The code is open-sourced on GitHub, meaning enterprise developers can extend this pattern to their own physical environments.
Temporal Intelligence: Knowing When the Job Is Done
One of the harder problems in physical AI has been task verification: how does a robot know when a step is complete before moving to the next one? Gemini Robotics ER 2 introduces what Google calls “temporal intelligence,” using continuous video understanding to track progress against specifications.
The practical outcome: a robot tightening a bolt or assembling a component can verify the step is done to specification before advancing, rather than moving on based on elapsed time or a fixed motion sequence. For enterprise use cases in quality control, assembly, and logistics, this closes a gap that has kept AI-driven physical automation confined to narrow, highly structured tasks.
Multi-Robot Collaboration: Fleet-Level Orchestration
Gemini Robotics ER 2 introduces multi-robot collaboration as a first-class capability. Multiple robots can share a workspace and divide complex workflows, completing tasks that a single robot physically cannot manage alone.
This is the physical-world equivalent of multi-agent AI systems, which enterprise teams have been adopting for software workflows over the past 18 months. The ARD (Agentic Resource Discovery) standard that Google and Microsoft co-developed for digital agent discovery finds a physical analogue here: fleets of specialized robots coordinating around shared task graphs.
The industrial applications are immediate: warehouse pick-and-pack operations requiring coordinated transport and manipulation, multi-station assembly lines, and hospital supply chains where different robot types handle different physical constraints. Google’s preliminary data shows multi-robot setups completing complex workflows faster than any single robot, with ER 2 handling the handoff sequencing automatically.
What Enterprise AI Teams Should Do Now
The immediate access point is Gemini Robotics ER 2 through the Gemini API. Enterprise teams that have already built agentic workflows using Gemini models can extend those same pipelines into physical environments by declaring robot control interfaces as tools.
The Gemini Enterprise Agent Platform private preview is the right place for organizations with existing Google Cloud relationships to begin structured evaluation. Early-access partnership for the VLA models is the path for companies with their own robot hardware.
Three actions worth taking this week:
Audit your physical workflows for agentic fit. The tasks where ER 2 adds most value are multi-step physical processes currently handled by teleoperation, narrow automation, or human labor. Logistics, manufacturing, and facilities management are the obvious categories, but any operation involving repeated physical manipulation in variable environments is a candidate.
Map your existing AI stack to the physical layer. If your team already uses Gemini models for digital agent workflows, the API surface for ER 2 is familiar. Developers who have built tool-calling pipelines for software agents can apply the same patterns to physical tools like VLA models and navigation APIs.
Watch the On-Device adaptation timeline. Gemini Robotics On-Device 2’s ability to adapt to a new robot embodiment in a few hours of data dramatically lowers the switching cost between hardware platforms. Enterprises evaluating robot hardware vendors no longer need to treat the AI model choice as locked to a specific chassis.
For enterprise teams still building the operational infrastructure to deploy AI agents at scale, the convergence of digital and physical agentic systems is accelerating faster than most roadmaps anticipated. The Gemini Enterprise Agent Platform is now the coordination layer across both.
If your team is evaluating where physical AI fits your AI transformation roadmap, talk to Enera about how to structure the assessment before the early-access windows close.
Key Takeaways
- Google DeepMind released Gemini Robotics 2 on July 30, 2026, comprising three models: a full-body VLA, an embodied reasoning orchestration layer, and an efficient on-device VLA.
- Gemini Robotics ER 2 is available now via the Gemini API, Google AI Studio, and in private preview on Gemini Enterprise Agent Platform.
- Multi-robot collaboration and temporal task verification are the two capabilities that close the gap between demo-grade and enterprise-grade physical AI.
- The tool-calling architecture mirrors existing enterprise agent patterns, meaning teams with Gemini API experience can extend to physical workflows without rebuilding their AI stack.
- The Gemini 3.6 Flash agentic stack and the robotics layer now share an API surface, giving enterprise teams a single integration path across digital and physical operations.