Deepmind Launches New Generation Robot AI Model: Gemini Robotics On-Device

AIbase基地

Published inAI News · 6 min read · Jun 25, 2025

Google DeepMind officially launched its next-generation robotic AI model, Gemini Robotics On-Device, marking a milestone in the development of robotic AI technology toward greater efficiency and independence. This model can run locally on robots without an internet connection and demonstrates strong versatility and task adaptability, bringing revolutionary breakthroughs to the fields of industrial, warehouse, and home service robots.

Breaking Cloud Limitations: Localized Operation of Robotic AI

Gemini Robotics On-Device is a vision-language-action (VLA) model based on Google's Gemini 2.0, with the key feature of running entirely on the robot's local hardware without relying on cloud computing resources. This solves issues of latency and reliability in traditional cloud-based robotic systems under unstable network conditions. Carolina Parada, Senior Director at DeepMind, said: "This model is compact and efficient, capable of running directly on robot hardware, ensuring stable performance with low latency and offline operation."

By operating locally, Gemini Robotics On-Device significantly enhances the practicality of robots in network-restricted scenarios such as factories, warehouses, or remote areas. Tests show that its performance is close to the cloud-based Gemini Robotics model, while surpassing other local AI models in multiple benchmark tests, demonstrating strong competitiveness.

Universality and Flexibility: From 50 Demonstrations to New Tasks

Gemini Robotics On-Device not only excels in performance but also stands out for its task adaptability. DeepMind claims that the model can quickly adapt to new tasks with just 50 to 100 demonstrations, such as unzipping a zipper, folding clothes, or performing industrial assembly.

The model was initially trained for the ALOHA robot, but it has successfully adapted to the dual-arm Franka FR3 robot and the Apollo humanoid robot from Apptronik, demonstrating its universality across different hardware platforms. Developers can control and fine-tune the model through natural language instructions, enabling it to easily handle complex dual-arm tasks or new objects in dynamic environments. Parada emphasized: "Generative AI allows robots to generalize from limited data, significantly accelerating deployment in complex scenarios."

Open Developer Ecosystem: SDK Empowers Innovation

To accelerate the industry application of Gemini Robotics On-Device, Google DeepMind has simultaneously released a software development kit (SDK), now open for applications through the "Trusted Tester" program on GitHub. Developers can use the SDK to test and fine-tune the model in Google's MuJoCo physical simulator or real-world environments. This move marks DeepMind's first time opening up the ability to fine-tune VLA models to developers, paving the way for customized applications of robotic AI.

The SDK enables developers to quickly train robots to perform specific tasks with just a few demonstrations, such as putting a Rubik's cube into a bag or handling delicate industrial operations. DeepMind stated that the model performs well in new scenes and objects it hasn't seen before, such as completing assembly tasks on an industrial conveyor belt, showing strong generalization capabilities.

Safety and Industry Prospects: The Next Step for Robotic AI

In terms of safety, DeepMind emphasized that Gemini Robotics On-Device includes comprehensive safety measures and collaboration with experts and policymakers to minimize potential risks. At the same time, the release of this model is seen as part of the fierce competition between Google and competitors like Nvidia GR00T and OpenAI RT-2 in the field of general robotic AI.

From warehouse robots to home service robots, the localized operation capability and rapid learning features of Gemini Robotics On-Device lay the foundation for its wide application in various scenarios. AIbase believes that this technology will not only reduce the cost of robot deployment but may also drive AI-driven automation into more daily life scenarios.

Model Access: https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/

Anthropic Launches New Feature: Users Can Directly Build AI Applications in Claude

Anthropic, an American startup focused on generative artificial intelligence, recently announced the launch of a new feature called "Artifacts," which allows users to create personalized applications. Users can create by means of simple conversations without any programming knowledge. This feature's release marks an important step for Anthropic in the field of artificial intelligence application development. The Artifacts feature was initially launched last June and became available to all users in August. Users can access it next to the conversation window.

New GoT-R1 Multimodal Model Released: Making AI Drawing Smarter, the New Era of Image Generation!

Recently, a research team from the University of Hong Kong, The Chinese University of Hong Kong, and SenseTime has released a groundbreaking framework - GoT-R1. This new multimodal large model significantly enhances the semantic and spatial reasoning capabilities of AI in visual generation tasks by introducing reinforcement learning (RL), successfully generating high-fidelity and semantically consistent images from complex text prompts. This advancement marks another leap in image generation technology. Currently, although existing multimodal large models have made significant progress in generating images based on text prompts

AI startup Scale AI exposed for leaking customer confidential information using Google Docs

Recently, the artificial intelligence startup Scale AI has been involved in a serious data security scandal. This highly valued company, which was acquired by Meta for 49% of its shares for 14.8 billion dollars, was exposed for using public Google Docs to store confidential information from numerous clients including Meta, Google, and xAI. Google Docs is a convenient collaboration tool, but its sharing options of either inviting or completely public access clearly conflict with any company's strict security standards. It was reported that anyone who knew Sca

Google Launches Gemini CLI! AI Assistant Touches the Developer Terminal

Recently, Google officially launched a new command-line tool - Gemini CLI. This tool is based on Google's self-developed Gemini2.5Pro AI model, aiming to provide developers with convenient AI Q&A and content generation services. With Gemini CLI, developers can directly call the power of AI in their own terminal interface, thereby improving programming efficiency and work convenience. One of the highlights of Gemini CLI is its support for up to 1 million tokens

ChatGPT iOS App Monthly Downloads Exceed 30 Million, Surpassing All Social Apps

The iOS app of ChatGPT had 29.6 million downloads in the past 28 days, becoming the most popular app globally. This achievement made ChatGPT's download count exceed the combined total of four social apps - TikTok, Facebook, Instagram, and X - which had approximately 32.9 million downloads during the same period, creating a difference of 10.6%. Although social apps have been on the market longer, ChatGPT achieved this in a short period of time.

Image Giant Getty Images Reverses Core Copyright Lawsuit Against Stability AI, UK Case Continues

Recently, Getty Images announced in the London High Court that it has withdrawn its main copyright infringement allegations against Stability AI, further narrowing the focus of this closely watched legal battle. The core of this lawsuit revolves around how AI companies use copyrighted content to train their models. Image source note: The image is AI-generated, and the image licensing service is Midjourney. Although Getty Images' dismissal of the case did not end it, the company is still pursuing other allegations.

Dou Bao AI's Gaokao Score Reaches the Threshold for Tsinghua and Peking University Admission! Literature Score of 683 Leads Domestic and International Top Models

ByteDance's Seed Team recently released the impressive results of the 2025 full subject Gaokao test: the Dou Bao Seed1.6-Thinking model scored 683 in literature and 648 in science, meeting the admission threshold for Tsinghua and Peking University, and performed outstandingly in domestic and international AI model Gaokao tests. The test used the national new volume one and Shandong Province's independent proposition papers, with Dou Bao competing against five other top domestic and international AI models such as Google Gemini 2.5 Pro, DeepSeek R1, and OpenAI o3.

Google Unveils a Major Move! Gemini CLI Open Source Release, Free AI Coding Assistant Challenges Cursor

Google has officially launched Gemini CLI today, an open-source terminal AI agent tool, directly challenging commercial AI coding tools. The project immediately received over 9,000 stars on GitHub, showing strong community interest. The free strategy disrupts the market: Zero cost usage: Simply use a personal Google account to obtain a Gemini Code Assist license. Top model: Free access to the Gemini 2.5 Pro model.