Google DeepMind Launches Gemini Robotics ER 2 and Humanoid Control Suite

On July 31, 2026, Google DeepMind officially launched Gemini Robotics ER 2, a major upgrade to its "embodied reasoning" model family designed to act as a high-level planner and cognitive brain for physical robots. Alongside ER 2, Google DeepMind unveiled a complete robotics control suite, including Gemini Robotics 2 (its first model to control a full humanoid under a single policy) and On-Device 2 (a lightweight model for local processing).

A Trio of Robotics Brains

The launch marks a transition from specialized, isolated robotic control systems to integrated, multi-model cognitive hierarchies¹:

"Google calls ER 2 “a high-level brain for robots” and shipped it alongside two siblings: Gemini Robotics 2, a vision-language-action model that controls full humanoids from feet to fingertips, and On-Device 2, a lightweight version that runs locally on the robot itself." — The Gadgeteer: Google’s Gemini Robotics ER 2 Wants to Run the Whole Robot Crew

Gemini Robotics ER 2 (High-Level Brain): Serves as the cognitive planning layer. It combines advanced video understanding, multi-step task orchestration, progress tracking, and self-correction. It achieves 91.3% accuracy on temporal "moment-finding" (understanding exactly when tasks begin/end and identifying key physical events). It is publicly available via the Gemini API and Google AI Studio.
Gemini Robotics 2 (Whole-Body Control): Google DeepMind's first AI model to control a full humanoid robot—coordinating legs, torso, arms, and multi-finger hands—under a single, end-to-end learned policy.
On-Device 2 (Local Execution): A lightweight version designed to run directly on robot hardware, ensuring low latency, safety, and operational continuity even without internet access.

Fluid Real-Time Orchestration

The primary advancement of ER 2 is its ability to reason in real-time without the awkward pauses that plagued prior generations:

"Gemini Robotics ER 2 represents a step change in powering robots with video understanding, task orchestration, and multi-robot collaboration — making it possible for robots to be more helpful in the physical world." — Google Blog: Introducing Gemini Robotics ER 2

By providing sub-second reasoning and spatial awareness comparable to massive frontier models like Claude Fable 5 or GPT-5.6 Sol, ER 2 enables fluid, real-time robot coordination. This includes multi-robot collaboration where independent machines can communicate, split tasks, and dynamically adapt to environmental changes.

An instance of Siloed AI applications are collapsing into unified conversational platforms. — Google DeepMind is replacing isolated, single-purpose robotic control mechanisms with an integrated multi-model cognitive hierarchy. ↩︎

Part of

This finding is an example of a pattern recurring across your work:

Industrial world data is the ultimate moat against internet-trained AI models.

Google DeepMind is launching models to power humanoid control and physical dexterity, shifting from digital reasoning to real-world physical environments.

Industrial world data is the ultimate moat against internet-trained AI models.

As internet-scraping matures, robotic laboratories capturing analog physical anomalies are building the ultimate data asset for physical foundation models.

Industrial world data is the ultimate moat against internet-trained AI models.

Apptronik's Robot Park aims to bypass the limitations of digital simulations by harvesting real-world physical training data from hundreds of robots.

Industrial world data is the ultimate moat against internet-trained AI models.

Apptronik's physical facility is designed specifically to capture hardware exceptions and real-world nuances that digital simulations completely miss.

Dual-use AI systems require physical testing environments to bridge simulation gaps.

Humanoid robotics platforms are building massive physical training hubs to capture real-world operational exceptions that digital simulations fail to model.

Revision history

Update the Google DeepMind humanoid robotics note to cover the official launch of Gemini Robotics ER 2, Gemini Robotics 2, and On-Device 2 on July 31, 2026, detail their respective roles, and highlight the 91.3% moment-finding accuracy.

33d ago· by the agent

Update the existing humanoid robotics note to cover the major product launch of Gemini Robotics 2 by Google DeepMind and its implementation on Apptronik's Apollo 2 robot.

36d ago· by the agent