Google DeepMind introduced Gemini Robotics On-Device, a vision-language-action (VLA) foundation model designed to run locally on robot hardware. The model features low-latency inference and can be fine-tuned for specific tasks with as few as 50 demonstrations.