Key Advancements in Gemini Robotics 2

1. Expansion from upper-body to full-body control

The previous Gemini Robotics model (March 2025) focused on controlling a robot's torso, arms, and hands. Gemini Robotics 2 extends control to the entire body — from feet to fingertips — enabling coordinated locomotion, posture, balance, bimanual manipulation, and walking . DeepMind describes this as "whole-body intelligence" that allows humanoid robots to navigate, reach, crouch, and recover from perturbations in a unified way .

2. The three-model suite

• Gemini Robotics 2 (VLA model): A Vision-Language-Action model that directly translates visual input and natural language instructions into motor commands for full-body humanoid control .
• Gemini Robotics ER 2 (Embodied Reasoning model): An advanced vision-language model focused on physical-world reasoning — spatial understanding, multi-step task planning, video comprehension, and orchestrating actions across multiple robots. It is available in public preview as of launch .
• Gemini Robotics On-Device 2 (Efficient Variant): A lightweight, optimized model designed to run directly on robot hardware. It can adapt to a new robot body using fewer than 200 training examples and a few hours of training, dramatically reducing deployment time .

3. Fine manipulation and multi-robot collaboration

• Fine dexterity: DeepMind reports 89.6% accuracy on the Franka Duo precision insertion benchmark, a task requiring two robot arms to align and insert objects with tight tolerances .
• Multi-robot collaboration: Multiple robots can now coordinate on shared tasks in shared spaces, with the model jointly reasoning about the positions and actions of other robots to avoid collisions and accomplish team objectives .

4. ASIMOV-Agentic safety benchmark

DeepMind released ASIMOV-Agentic, an open-source safety benchmark (available on Hugging Face under CC-BY-4.0) designed to evaluate whether a robot's commanded actions could lead to physically harmful outcomes . It builds on DeepMind's earlier ASIMOV work (grounded in real-world injury reports) and introduces agentic scenarios where models must detect when a person is nearby, halt movement until the area is clear, and only then resume . Gemini Robotics ER 2 is described as DeepMind's safest model yet, with multi-layered guardrails across every model layer .

Industry Context: FCC Ban on Chinese-Made Robots

On July 28, 2026, the FCC added "foreign-produced advanced robotic devices" — including humanoid robots and quadruped (four-legged) robots — to its Covered List, effectively blocking new imports, marketing, and sales in the U.S. . The ban also covers connected power inverters. The administration cited national security risks, arguing that Chinese-made robots could create supply-chain vulnerabilities, be exploited to threaten critical infrastructure, and pose cybersecurity risks . The order targets China, the dominant global producer of such devices, but does not affect previously purchased robots . China has publicly opposed the ban .

Timing significance: The FCC ban and DeepMind's Gemini Robotics 2 launch occurred within the same week — the ban removes a major source of foreign competition in the U.S. humanoid robot market, while DeepMind's whole-body AI platform positions U.S.-aligned robotics firms (DeepMind is partnering with Apptronik, Boston Dynamics, and others ) to fill the gap with domestically developed, AI-native hardware .

Key Advancements in Gemini Robotics 2

1. Expansion from upper-body to full-body control

The previous Gemini Robotics model (March 2025) focused on controlling a robot's torso, arms, and hands. Gemini Robotics 2 extends control to the entire body — from feet to fingertips — enabling coordinated locomotion, posture, balance, bimanual manipulation, and walking . DeepMind describes this as "whole-body intelligence" that allows humanoid robots to navigate, reach, crouch, and recover from perturbations in a unified way .

2. The three-model suite

• Gemini Robotics 2 (VLA model): A Vision-Language-Action model that directly translates visual input and natural language instructions into motor commands for full-body humanoid control .
• Gemini Robotics ER 2 (Embodied Reasoning model): An advanced vision-language model focused on physical-world reasoning — spatial understanding, multi-step task planning, video comprehension, and orchestrating actions across multiple robots. It is available in public preview as of launch .
• Gemini Robotics On-Device 2 (Efficient Variant): A lightweight, optimized model designed to run directly on robot hardware. It can adapt to a new robot body using fewer than 200 training examples and a few hours of training, dramatically reducing deployment time .

3. Fine manipulation and multi-robot collaboration

• Fine dexterity: DeepMind reports 89.6% accuracy on the Franka Duo precision insertion benchmark, a task requiring two robot arms to align and insert objects with tight tolerances .
• Multi-robot collaboration: Multiple robots can now coordinate on shared tasks in shared spaces, with the model jointly reasoning about the positions and actions of other robots to avoid collisions and accomplish team objectives .

4. ASIMOV-Agentic safety benchmark

DeepMind released ASIMOV-Agentic, an open-source safety benchmark (available on Hugging Face under CC-BY-4.0) designed to evaluate whether a robot's commanded actions could lead to physically harmful outcomes . It builds on DeepMind's earlier ASIMOV work (grounded in real-world injury reports) and introduces agentic scenarios where models must detect when a person is nearby, halt movement until the area is clear, and only then resume . Gemini Robotics ER 2 is described as DeepMind's safest model yet, with multi-layered guardrails across every model layer .

Industry Context: FCC Ban on Chinese-Made Robots

On July 28, 2026, the FCC added "foreign-produced advanced robotic devices" — including humanoid robots and quadruped (four-legged) robots — to its Covered List, effectively blocking new imports, marketing, and sales in the U.S. . The ban also covers connected power inverters. The administration cited national security risks, arguing that Chinese-made robots could create supply-chain vulnerabilities, be exploited to threaten critical infrastructure, and pose cybersecurity risks . The order targets China, the dominant global producer of such devices, but does not affect previously purchased robots . China has publicly opposed the ban .

Timing significance: The FCC ban and DeepMind's Gemini Robotics 2 launch occurred within the same week — the ban removes a major source of foreign competition in the U.S. humanoid robot market, while DeepMind's whole-body AI platform positions U.S.-aligned robotics firms (DeepMind is partnering with Apptronik, Boston Dynamics, and others ) to fill the gap with domestically developed, AI-native hardware .