Google rolls out new Gemini model that can run on robots locally
Google DeepMind has unveiled its latest breakthrough in artificial intelligence: the Gemini Robotics On-Device model. This innovative language model empowers robots to execute complex tasks locally, eliminating the need for a constant internet connection. This development marks a significant step forward in autonomous robotics, promising greater efficiency and reliability for AI-driven systems.
Building upon the foundation of the previous Gemini Robotics model, which debuted in March 2025, the new On-Device iteration specializes in controlling a robot’s physical movements. Developers now have the unprecedented flexibility to manage and fine-tune the model’s behavior using intuitive natural language prompts, simplifying the programming process and accelerating deployment across various applications.
In rigorous benchmarks, Google reports that the Gemini Robotics On-Device model achieves performance levels remarkably close to its cloud-based predecessor. Furthermore, it claims to outperform other on-device models in general benchmarks, though specific competitors were not named. This suggests a significant leap in local processing capabilities for robotics AI.
Demonstrations highlighted the model’s practical capabilities, with robots powered by Gemini Robotics On-Device performing intricate tasks such as unzipping bags and folding clothes. While initially trained for ALOHA robots, Google successfully adapted the model to operate on other advanced platforms, including a bi-arm Franka FR3 robot and Apptronik’s Apollo humanoid robot. Notably, the bi-arm Franka FR3 showcased its prowess by efficiently tackling scenarios and objects it had not previously encountered, such as complex industrial assembly tasks on a belt.
To further empower the robotics community, Google DeepMind is also releasing a Gemini Robotics SDK. This software development kit will enable developers to train robots on new tasks by providing 50 to 100 demonstrations within the MuJoCo physics simulator, accelerating the development and deployment of new robotic functionalities.
Google’s strides in robotics AI are part of a broader industry trend. Major players like Nvidia are developing foundational models for humanoids with platforms such as GR00t, while Hugging Face is not only creating open models and datasets but also actively developing humanoid robots. Additionally, Korean startup RLWRLD, backed by Mirae Asset, is focusing on foundational models for robotics. This surge of innovation underscores a collective vision for a future where intelligent, autonomous robots play an increasingly vital role in various sectors.
