Google DeepMind Launches Gemini Robotics 2 Model

Google DeepMind has unveiled Gemini Robotics 2, its latest vision-language-action model designed for physical systems. The system introduces what the company calls "intelligent whole-body control," allowing humanoid robots to perceive their environments, reason through multi-step procedures, and coordinate movements across their entire structures. According to Carolina Parada, the robotics lead at DeepMind, the ultimate goal of this technology is to empower physical machines to execute any task a human can perform.

Recent video demonstrations showcased the model's capabilities using Apptronik's Apollo 2 robot and robotic hands developed by Sharpa. The machines autonomously completed everyday household and industrial chores, such as screwing in lightbulbs and tying up trash bags. To achieve these results, the system was trained using a combination of simulation, video demonstrations, and teleoperation. The model also enables different robots to collaborate with one another on shared tasks.

For robotics practitioners, this release marks a shift from programming isolated components to managing unified systems. Forrester analyst Paul Miller noted that Gemini Robotics 2 improves on previous iterations by helping machines recover from failed actions and fostering better coordination between multiple robots. Instead of focusing strictly on individual parts like a gripper or an arm, the model treats the robot as a single, integrated entity.

Despite these advancements, significant hurdles remain before these systems can be deployed alongside humans. Miller emphasized that developers must still prove the safety of these heavy physical machines, particularly regarding how they fail when power systems or sensors malfunction. Consequently, practitioners are expected to deploy this physical AI first in highly controlled environments like warehouses and factories, where risks and uncertainties can be minimized, before attempting domestic deployments.