Google DeepMind has unveiled a new AI model that allows robots to operate with full autonomy, untethered from the cloud. The new system, named Gemini Robotics On-Device, runs entirely on a robot’s local hardware, a critical development that promises to make robotic systems faster, more reliable, and capable of functioning in environments with intermittent or no internet connectivity. This marks a significant step toward making general-purpose robots practical for real-world applications.
Google DeepMind explains that the new model, designed for bi-arm robots, is not just efficient but also highly adaptable and engineered to require minimal computational resources. It is the first of the company’s vision-language-action (VLA) models to be made available for fine-tuning, allowing developers to adapt it for new, highly dexterous tasks with as few as 50 to 100 demonstrations. This ability to generalize from a small amount of new data could dramatically accelerate the deployment of robots in complex settings.
To facilitate this, the company is releasing a Gemini Robotics SDK, available on GitHub, through a selective trusted tester program, which developers can apply for via Google. This controlled rollout underscores the model’s strategic importance as the race to build more capable physical AI systems intensifies among major tech labs.
The Race to Cut the Cord: On-Device vs. Cloud AI
Google’s move is a major entry in a pivotal debate shaping the future of robotics: whether intelligence should live in the cloud or directly on the machine. On-device processing is crucial for real-time robotics because it eliminates the network latency inherent in cloud computing. For robots interacting with the physical world, a split-second delay in decision-making can be the difference between success and failure. This makes local AI essential for applications where instant responses are non-negotiable.
The main trade-off for on-device AI, however, is the inherent limitation of local hardware, which has less computing power and storage than vast cloud servers. This challenge has led to different strategic bets across the industry. Figure AI, for instance, introduced its robotics optimized Helix AI model in February, which, like Google’s new system, runs entirely on embedded GPUs.
In contrast, Microsoft’s Magma AI model is designed for deep integration with its Azure cloud platform, targeting enterprise automation where connectivity is more reliable. Google’s own strategy has evolved; its flagship Gemini Robotics platform, introduced in March, uses a hybrid approach. The new on-device model provides a dedicated solution for scenarios where autonomy is paramount.
A Crowded Field of Physical Intelligence
Venture capital investment in industrial humanoid robotics tripled in 2024 to $1.2 billion, signaling intense competition. According to the International Federation of Robotics, the global market for industrial robot installations has already hit an all-time high of $16.5 billion, with a key trend for 2025 being “Physical AI”—systems that learn from experience rather than rigid programming.
This competitive landscape includes major players pursuing unique philosophies. While Google and Figure AI champion on-device speed, Meta recently released V-JEPA 2, an open-source “world model” that learns physical common sense from video. These models allow an AI to run internal simulations to “think” before it acts, letting machines “plan movements and interactions in simulated spaces” before attempting them in the physical world.
This approach dramatically reduces costly trial-and-error and accelerates learning for tasks in industrial assembly and logistics. This method, focused on building an internal understanding of physics, offers another path toward creating robots that can navigate unpredictable human environments.
Teaching Robots to Learn Like Humans
At the heart of Google’s new model is a focus on generalization—the ability to perform new tasks with minimal training. This is achieved through a technique known as Few-shot learning (FSL), which allows a model to learn from a very small number of examples. This approach seeks to emulate the human ability to grasp new concepts quickly, a stark contrast to traditional AI models that often require millions of data points. For robotics, where collecting vast, labeled datasets for every possible task is impractical, FSL is a game-changer.
Google claims Gemini Robotics On-Device can be adapted with as few as 50 to 100 demonstrations. The company provided concrete evidence of this adaptability, noting that while the model was initially trained for ALOHA robots, it was successfully adapted to a bi-arm Franka FR3 robot and the Apollo humanoid robot by Apptronik.
This capability is what enables the system’s broader potential. As Carolina Parada, head of robotics at Google DeepMind, explained in reporting from Ars Technica, the model’s generative power extends beyond simple commands. “It’s drawing from Gemini’s multimodal world understanding in order to do a completely new task… What that enables is in that same way Gemini can produce text, write poetry, just summarize an article, you can also write code, and you can also generate images. It also can generate robot actions.”
From Open Science to Guarded Advantage
Google’s decision to release its new robotics SDK through a limited program highlights a broader strategic pivot within DeepMind. The lab, once a bastion of open scientific publication, now more selectively releases its core technology to protect Google’s competitive advantage. This shift has reportedly caused friction internally, with one researcher telling the Financial Times, “We are now told that publication is no longer the default.”
This proprietary stance contrasts sharply with Meta’s role in open-source AI with its Llama models, a strategy designed to accelerate community innovation. While this openness is lauded, the performance of open models has historically trailed their closed-source counterparts. The best open-source models have lagged behind proprietary ones by several months, though that gap is shrinking. This performance difference helps explain why a company like Google would guard its most advanced technology, even as it provides tools for developers to build upon it.
Google’s release of Gemini Robotics On-Device is a calculated move in the high-stakes competition to build the next generation of intelligent machines. It directly addresses the industry’s critical need for low-latency, autonomous systems while showcasing remarkable advances in rapid, human-like learning. Yet, the model’s ultimate impact will be shaped not only by its technical prowess but also by the strategic tension between the collaborative spirit of open research and the guarded realities of commercial competition.
