Google DeepMind has unveiled a significant upgrade to its Gemini Robotics AI model, allowing humanoid robots to perform a wider range of actions from walking and crouching to picking up objects. The new version, Gemini Robotics 2, can control entire humanoid robots, from their feet to fingertips, enabling them to bend over and pick up items like watering cans or specific objects off shelves.

Though the model still has room for improvement in terms of movement speed, it represents an important step towards more complex, real-world tasks. The updated model also supports better dexterity, allowing robots to perform intricate tasks such as sealing Ziploc bags and unscrewing lightbulbs with five-fingered hands.

DeepMind is simultaneously updating its Gemini Robotics ER (embodied reasoning) vision-language model, which helps robots analyze their surroundings and process instructions. The new version can now complete tasks over an extended period of time, better understand when tasks begin and end, and even coordinate multiple types of robots to work together.

However, DeepMind emphasizes that Gemini Robotics ER 2 is its safest robotics model yet, featuring enhanced detection for nearby humans and the ability to trigger safety calls if someone approaches too closely. Meanwhile, improvements have been made to the Gemini Robotics On-Device Model to enable faster adaptation to new robot embodiments with drastically different shapes, sensors, and degrees of freedom.