Google DeepMind has released Gemini Robotics 2, a new AI model that can control a range of robots, including humanoids capable of performing dextrous tasks such as screwing in lightbulbs and tying trash bags. The system integrates several AI models into a single framework, allowing robots to interpret their environment and execute actions. A vision language model (VLM) enables communication with humans and reasoning about tasks, while two vision language action (VLA) models manage physical movement and hand operations. In video demos, the company showcased robots performing complex tasks autonomously using the combined model. One example featured Apptronik’s Apollo 2 robot using Sharpa hands to tidy shelves. The model was trained using a mix of human teleoperation, video examples, and simulations, though it still requires specific training for complex tasks. Google DeepMind emphasized that AI models cannot yet perform a wide range of tasks without such training. The release highlights Google’s growing focus on applying AI beyond digital environments to achieve its full potential. The company has a history of robotics research and previously partnered with Boston Dynamics to develop legged robots. 'It's another milestone in our path towards really getting towards what we call like physical AGI, which means we get a robot to do anything that a human can,' said Carolina Parada, head of robotics at Google DeepMind.
Google DeepMind's approach to safety involves multi-layered guardrails across models and the introduction of ASIMOV-Agentic, a new benchmark for evaluating the safety of AI systems controlling robots. The benchmark assesses whether commands could lead to harmful or uncertain outcomes. Parada noted that the safety question becomes more critical as AI models are deployed in real-world settings, where unpredictable behaviors may emerge. 'There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply,' she said. The company’s CEO, Demis Hassabis, previously expressed hopes for developing an AI operating system for robots, akin to Android for smartphones.
Source: wired
