Google DeepMind says new Gemini model can control humanoid robots

In short: Google DeepMind released Gemini Robotics 2, a new version of Gemini that can help control several kinds of robots, including humanoid robots.

What happened

Google DeepMind says its new system, called Gemini Robotics 2, can control robots that do hands-on tasks. The company says this includes humanoid robots, which are robots shaped like people, with arms and hands.

Gemini Robotics 2 combines multiple AI models into one setup. One part is a vision language model, which is an AI that can understand what it sees in images or video and also understand and produce words (like a robot’s eyes and a translator working together). Two other parts help the robot decide how to move its body and hands in real space.

In videos shared ahead of the release, Google DeepMind showed different robots doing tasks on their own. One example used Apptronik’s Apollo 2 robot with hands from a company called Sharpa to tidy shelves. DeepMind said it trained the system using a mix of human-guided practice, video examples, and simulations, which are computer-made practice environments.

Why it matters

AI that can move through the real world can be useful in places like warehouses, factories, and possibly homes. But it also raises safety concerns, because a robot can cause real damage if it does something unexpected. DeepMind says it uses “guardrails” at different levels and it is introducing a new test called ASIMOV-Agentic to check whether a command could lead to harmful or uncertain results.

Source: Wired