Google DeepMind Unveils Gemini Robotics 2 for Whole-Body Humanoid Control
The new vision-language-action model enables humanoid robots to coordinate movements from feet to fingertips for complex real-world tasks.
Google DeepMind has introduced Gemini Robotics 2, a vision-language-action (VLA) model designed to provide whole-body control for humanoid robots. This development marks a significant shift in robotic AI, moving beyond isolated limb movements to a system that coordinates a robot's entire physical form.
The system is deployed as a suite of specialized models. The core VLA handles primary motion, while Gemini Robotics ER 2 manages high-level embodied reasoning and collaboration between multiple robots. For local execution, Gemini Robotics On-Device 2 allows for faster processing and adaptation. According to Google DeepMind, the system manages the entire robot body, enabling humanoids to walk, crouch, and manipulate objects simultaneously.
From Tabletops to Total Control
Previously, DeepMind's robotics efforts were largely confined to tabletop environments and upper-body tasks. These earlier iterations lacked the integration necessary for a robot to navigate a room while performing a manual task. The transition to whole-body coordination is intended to move humanoid utility into complex, real-world environments where navigation and manipulation must occur in tandem to be effective.
To demonstrate these capabilities, the model was tested on the Apptronik Apollo 2 humanoid. The robot utilized a SharpaWave five-fingered hand featuring 22 degrees of freedom to perform high-precision tasks, including sealing ziplock bags and tying knots. These demonstrations highlight the model's ability to translate high-level AI instructions into the granular motor skills required for delicate physical work.
Implications for General Purpose Robotics
This advancement bridges a critical gap between high-level AI reasoning and low-level physical execution. By integrating embodied reasoning with whole-body VLA, robots can now manage multi-step tasks that last several minutes and operate more safely around humans. This integration is a necessary step toward the viable deployment of general-purpose robots in workplaces and residential homes, where they must interact with unpredictable physical spaces.
The Path to Deployment
As the industry moves toward more capable humanoids, the focus shifts to how these models scale across different hardware. While the Apptronik Apollo 2 served as the primary demonstration platform, the modular nature of the Gemini Robotics 2 suite suggests a goal of broader hardware compatibility. Future developments will likely focus on refining the collaboration between multi-robot systems and increasing the autonomy of on-device execution in diverse environments.