In context
In mid-2025, the robotics industry was pushing toward more autonomous and responsive systems, but many AI-driven robots still relied on cloud connectivity, which introduced latency and reliability issues. Google DeepMind's release of an on-device vision-language-action (VLA) model addressed this gap, aiming to embed advanced reasoning directly into local robotic hardware.
What was reported
Google DeepMind introduced Gemini Robotics On-Device, an optimized version of its Gemini Robotics VLA model launched in March. Designed to run locally on robots, it ensures robust performance in environments with limited or no network connectivity and provides low-latency inference critical for sensitive applications.
The model supports bi-arm robots and demonstrates strong general-purpose dexterity and task generalization. It can follow natural language instructions and execute tasks like unzipping bags or folding clothes. Evaluations show it outperforms previous on-device models in challenging out-of-distribution tasks and complex multi-step instructions.
DeepMind also released a Gemini Robotics SDK for testing in the MuJoCo simulator, enabling rapid adaptation to new domains with as few as 50 to 100 demonstrations. The model is the first VLA from DeepMind available for fine-tuning and has been adapted to various embodiments, including the Aloha, Franka FR3, and Apptronik's Apollo humanoid.
Safety measures include semantic and physical safety layers, with recommendations for red-teaming and use of semantic safety benchmarks. Initial access is limited to trusted testers.
Why it mattered
This development signaled a shift toward edge AI in robotics, reducing reliance on cloud infrastructure and enabling faster, more reliable operation in real-world manufacturing and service settings. The ability to fine-tune with minimal demonstrations could accelerate deployment across diverse tasks and robot platforms, making advanced automation more accessible.
“Gemini Robotics On-Device is specifically engineered to operate locally on robotic devices, ensuring robust performance in environments with limited or no network connectivity.”
Source: Robotics & Automation News (roboticsandautomationnews.com) · Published 2025-06-26 · “Google DeepMind launches new vision language action model to ‘put AI directly into local robotic devices’”