Google DeepMind's new Gemini Robotics 2 model moves beyond tabletop tasks to enable full-body coordination and dexterity in humanoid robots.
Google DeepMind has moved its vision-language-action (VLA) models from simple tabletop manipulation to complex, coordinated physical movement. The company recently unveiled Gemini Robotics 2, an update designed to grant humanoid robots control over their entire physical structure rather than just their upper bodies.
This new iteration focuses on intelligent whole-body control and advanced dexterity. While previous versions were limited to localized tasks, the updated model allows a robot to reason through a sequence of movements. This includes walking, crouching, and stretching to complete multi-step objectives in unstructured environments.
One primary demonstration involved Apptronik's Apollo 2 humanoid. Using the SharpaWave hand, which features five fingers and 22 degrees of freedom, the robot successfully executed complex instructions. For instance, when told to place a watering can in a specific bin on a bottom shelf, the system managed the transition from walking to grasping and precise placement.
Integration and deployment
DeepMind claims the model can run locally on-device, a critical requirement for reducing latency in real-world robotics. Furthermore, the system is designed for rapid adaptation, reportedly allowing it to adjust to entirely new robotic hardware in only a few hours. This flexibility is a significant departure from traditional robotics, where new hardware often requires extensive retraining.
Beyond individual movement, the update introduces multi-robot collaboration. This capability allows multiple units to coordinate their actions to finish a task more efficiently. According to The Robot Report, the model translates high-level intent into the physical coordination required for such teamwork.
Access to these tools is currently being rolled out in stages. Gemini Robotics ER 2 is now available via Google AI Studio and is in private preview on the Gemini Enterprise Agent Platform. Meanwhile, the specific VLA and on-device models are being provided to early-access partners for testing.
The broader landscape
The shift toward embodied artificial intelligence comes as the industry faces intense pressure to prove commercial utility. While much of the current focus remains on generative text and images, the race to integrate these models into physical hardware is accelerating. This transition represents a move from digital assistants to physical agents capable of navigating human spaces.
This development occurs against a backdrop of massive capital shifts in the sector. As companies like OpenAI prepare for potential public market debuts, the competition to dominate the enterprise and industrial sectors is intensifying. The ability to move from software-based productivity to physical automation could define the next phase of the industry.
DeepMind admits that movement speed remains a hurdle for these systems. However, the leap from upper-body manipulation to whole-body coordination suggests that the foundational logic for complex, real-world labor is being established. The question remains whether these models can scale from controlled demonstrations to the unpredictable chaos of a standard warehouse or home.
FAQ
What is Gemini Robotics 2?
It is a vision-language-action model from Google DeepMind that enables humanoid robots to perform whole-body movements and complex tasks through reasoning.
Which robots can use this model?
DeepMind demonstrated the model using Apptronik's Apollo 2 humanoid equipped with a SharpaWave hand.
Can the model run without an internet connection?
Yes, DeepMind states that the model can run locally on-device to facilitate real-time robotic control.
How does it differ from previous versions?
Earlier versions focused on upper-body tasks like tabletop manipulation, whereas Gemini Robotics 2 enables full-body coordination, including walking and crouching.
