Google DeepMind has introduced Gemini Robotics On-Device, an advanced robotics model optimized to run locally on robotic hardware. This VLA (vision language action) model brings multimodal reasoning and real-world understanding directly to physical devices, operating independently of network connections. The on-device approach reduces latency and ensures robust performance in environments with intermittent connectivity.

The model demonstrates strong general-purpose dexterity across tasks like unzipping bags and folding clothes. Developers can access the Gemini Robotics SDK to evaluate and adapt the model with as few as 50-100 demonstrations. Built for bi-arm robots, it requires minimal computational resources while outperforming previous on-device alternatives on complex multi-step instructions and out-of-distribution tasks.

Gemini Robotics On-Device marks the first VLA model available for fine-tuning, with successful adaptation to different robot embodiments including Franka FR3 and Apollo humanoid robots. The model follows natural language instructions and manipulates objects, even previously unseen ones. Google is releasing it initially to a trusted tester program while applying holistic safety approaches aligned with its AI Principles.