Google Bets on Gemini Robotics as Universal Brain

In an interview with The Information, DeepMind’s CEO outlines plans to supply AI models to partner manufacturers, mimicking Android’s platform model against Tesla’s vertical approach.

Google DeepMind wants to become the operating system for robotics rather than building proprietary humanoid hardware. In an interview with The Information published in early October 2026, DeepMind CEO Koray Kavukcuoglu confirmed the division’s priority: supplying its Gemini Robotics model family to machinery manufacturers. The strategy brings Google’s classic Android playbook into the physical world, distributing software at scale while leaving assembly lines and factory capital to established hardware makers.

The Android playbook versus Tesla’s closed stack

That choice puts Google on a collision course with Tesla. While Elon Musk’s company designs both the Optimus chassis and its internal computer vision stack, Google is positioning itself as the default brain for third-party industrial arms and bipedal machines. The dynamic mirrors tech’s classic platform rivalry: Apple’s closed hardware-software integration against Android’s open, horizontal ecosystem.

Industry reports indicate that Tesla accelerated Optimus assembly in Fremont to hundreds of units per week by late September. Despite that production ramp, the automaker still grapples with software generalization when robots must handle unstructured tasks without manual intervention. Google aims directly at that intelligence gap, without carrying the balance-sheet burden of metal and motors.

Inside the Gemini Robotics 2 architecture

Google’s technical foundation rests on the Gemini Robotics 2 family, introduced by DeepMind in late July 2026. The suite is structured around three complementary modules:

Core Gemini Robotics 2: a whole-body Vision-Language-Action (VLA) system coordinating movement from feet to fingertips.
Gemini Robotics ER 2: built for embodied reasoning, multi-step task planning, and coordinating multiple robots in parallel.
On-Device 2: a lightweight variant running local execution directly on the machine’s onboard hardware.

According to DeepMind, the On-Device variant can adapt to an unfamiliar physical chassis within hours, requiring fewer than 200 recorded demonstrations to calibrate motor trajectories and grasping patterns.

Field partnerships and distribution

The software has already moved past pure benchmark simulations. Google demonstrated the system driving Apptronik’s Apollo 2 humanoid, including controlling mechanical hands with 22 degrees of freedom. The company also maintains active testing partnerships with Boston Dynamics, Agile Robots, and Franka.

For developers, the ER 2 reasoning model is accessible through Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, where more than 100 enterprise testers are evaluating the system.

Real-world hurdles and physical telemetry

Bypassing hardware manufacturing protects Google from the capital drains of actuator supply chains, custom alloys, and mechanical wear that routinely stall robotics startups. At the same time, deploying software across varied third-party machines generates diverse physical interaction data from real environments, an essential feedback loop for training future model iterations.

Significant questions remain unresolved. Google has not published commercial licensing terms for mass production, and DeepMind engineers concede that walking speed and fine object manipulation still require refinement before robots can work reliably in unconstrained environments.