谷歌Gemini 4 Argon：整合型機器人控制模型

google has unveiled gemini 4 argon, a single ai model intended for robotics control, simulation, and training data evaluation. argon can generate simulations from simple inputs like photos or sketches, allowing robots to practice movements before real-world deployment. it has been demonstrated controlling a boston dynamics spot robot to pick up cans and toss them into a trash bin. argon also works with other sources, such as recordings of teleoperated robots, first-person videos of humans performing tasks, and even ballpoint pen sketches. the model has shown improvements in success rates and execution times for robot tasks. this development is another step towards 'physical agi,' a goal google and deepmind have been pursuing.