Google DeepMind Unveils Gemini Robotics 2, Pushing Physical AI to Whole-Body Control

A humanoid robot named Apollo 2 can now unscrew a lightbulb with 92% reliability, but it struggles to sweep a floor with a dustpan, succeeding only 32% of the time. That gap in performance illustrates the bleeding edge of physical artificial intelligence. On July 30, Google DeepMind released Gemini Robotics 2, a suite of three models designed to serve as a universal intelligence layer for machines. The release marks a definitive shift from tabletop robotic arms to whole-body humanoid control, exposing both the rapid progress and the persistent physical limits of modern robotics.[1][2]

Historically, industrial robotics has been strictly hardware-centric. A robotic arm is engineered and programmed for a single workstation, executing a narrow, repetitive sequence of movements. If the environment changes, the robot fails. Gemini Robotics 2 attempts to rewrite this "robot assumption" by decoupling the intelligence from the machine. Instead of embedding logic within one tightly defined chassis, the system perceives an environment, interprets plain-language instructions, reasons about the task, and translates those decisions into joint movements across entirely different robot bodies.

The architecture relies on three distinct models working in concert. The first, Gemini Robotics 2, is a vision-language-action (VLA) model. It ingests camera pixels and text prompts, outputting direct motor control commands. Unlike previous iterations that only managed upper-body tabletop tasks, this VLA coordinates locomotion and manipulation simultaneously. It allows a humanoid to walk, reposition its center of gravity, crouch, reach, and handle an object as a single, fluid action.[1]

The second model, Gemini Robotics ER 2, acts as the system's high-level brain. "ER" stands for embodied reasoning. Built on the Gemini 3.5 Flash architecture, this vision-language model does not directly move the robot's joints. Instead, it orchestrates. It communicates with humans, interprets the physical space, and constructs multi-step plans that take several minutes to execute. It tracks progress, recognizes when a sub-task is complete, and can divide labor by coordinating multiple robots in a shared workspace.[1][2]

The third component, Gemini Robotics On-Device 2, addresses the latency and connectivity constraints of real-world industrial environments. It is a lightweight VLA optimized to run locally on the robot's own hardware. DeepMind reports that this on-device model can adapt to a previously unseen dual-arm robot using fewer than 200 demonstration examples, requiring only a few hours of training to map its outputs to the new machine's specific joints and sensors.[1]

The data on the system's physical dexterity reveals exactly where the technology currently excels and where it falters. The VLA model successfully controls the 22-degree-of-freedom SharpaWave hand on the Apptronik Apollo 2 humanoid, executing delicate tasks like tying knots or sealing a ziplock bag. DeepMind demonstrated a single model checkpoint driving the Apollo 2 with two different hand configurations, as well as a separate Franka Duo platform equipped with a standard two-fingered parallel gripper.[1][2]

Yet multi-finger dexterity remains the weakest axis in physical AI. While the system achieves a 92% success rate on highly structured tasks like unscrewing a bulb, performance drops precipitously on tasks requiring complex, continuous contact with deformable objects or tools. Sweeping with a dustpan—a task that demands constant micro-adjustments to maintain flush contact with the floor while coordinating two arms—succeeds only 32% of the time. The evidence suggests that while the cognitive reasoning is advancing rapidly, the physical translation of those thoughts into high-friction, dynamic environments is still a bottleneck.[2]

To manage the inherent risks of autonomous machines operating near humans, DeepMind introduced ASIMOV-Agentic, a new safety benchmark released under an open-source license. The ER 2 model is designed to evaluate its own uncertainty. If a user requests an action that violates physical constraints—such as handling liquids with an incompatible gripper, or lifting an object exceeding the robot's payload capacity—the model is trained to refuse the tool call and proactively request human intervention.[2]

The commercial rollout reflects the experimental nature of the hardware. Gemini Robotics ER 2 is available in public preview for developers via Google AI Studio, allowing software engineers to test the reasoning engine. However, the VLA and On-Device models remain strictly gated to early-access hardware partners, including Boston Dynamics, Apptronik, and Agile Robots. This tiered access prevents developers from deploying unverified motor-control models onto consumer hardware.[1][2]

DeepMind's strategy is clear: the company is not trying to build the ultimate humanoid robot. Instead, it is competing to build the operating system that will run on everyone else's hardware. By proving that a single intelligence layer can drive a Boston Dynamics Atlas, an Apptronik Apollo, and a Franka robotic arm, Google is positioning Gemini Robotics as the foundational infrastructure for the impending wave of physical AI.