On July 30, 2026, Google DeepMind announced Gemini Robotics 2, an artificial intelligence model designed to control the entire body of a humanoid robot. Unlike its predecessors, which were limited to arms or hands, the new system also controls lower limbs, torso, and five-fingered hands. The robots can walk, crouch, stretch, and perform fine manipulation tasks: sealing envelopes, unscrewing light bulbs, grasping irregularly shaped objects.
A Model for Embodied Reasoning and Collaboration
Alongside Gemini Robotics 2, DeepMind also released Gemini Robotics ER 2, a vision-language model dedicated to “embodied reasoning”. ER 2 can recognize the start and end of multi-step activities and orchestrate collaboration among different types of robots. A mobile robot can transport a component and hand it over to a fixed robot with a gripper to complete an assembly.
The on-device model (Gemini Robotics On-Device Model) operates without connectivity and quickly adapts to different morphologies, from Google’s dual-arm robots to the Apptronik Apollo 2 humanoid. This broadens potential use cases: logistics, assistance, and offline environments.
More Autonomy, More Risk
Extending control to the entire body and offline operation increase human-machine proximity. Kaushik Subramanian, head of robotics safety at DeepMind, called Gemini Robotics ER 2 “the safest model ever built.” It includes a human presence detection system that triggers an emergency stop if a person gets too close.
On-device execution reduces exposure to network attacks but doesn’t eliminate the problem. A compromise of the firmware or the model could disable safety systems. And the ability to control a robot’s entire body in shared spaces turns a cyberattack into a real physical risk.
Collaboration among heterogeneous robots introduces operational interdependencies. A single compromised node could propagate incorrect instructions throughout the entire system. No CVEs have been disclosed, but the attack surface grows with each new degree of freedom.
Mitigations Are Just a Starting Point
DeepMind hasn’t described specific hardening measures beyond proximity detection and offline operation. General principles for autonomous systems still apply: network segmentation, strong authentication for firmware updates, continuous validation of model integrity. Deploying humanoids in production and assistance environments will require stricter standards. And a regulatory framework that currently doesn’t exist.




