Google DeepMind unveils Gemini Robotics 2, enabling humanoid robots to perform complex chores

Google DeepMind released Gemini Robotics 2, a composite AI architecture that equips humanoid robots with the ability to execute dexterous actions such as installing lightbulbs and securing trash bags. The platform integrates a vision-language model for visual understanding and communication, plus two vision-language-action models that govern both whole-body locomotion and precise hand movements. In video previews, the Apollo 2 robot from Apptronik, fitted with Sharpa’s manipulators, organized shelves without human input, after being trained via teleoperation, video data, and simulated environments.

While the breakthrough showcases broader robotic capability, Google acknowledges the safety challenges of deploying frontier AI in physical settings and has introduced multi-layered safeguards and the ASIMOV-Agentic safety benchmark. The move follows prior collaborations with Boston Dynamics and signals the company’s ambition to create a universal AI operating system for robots, akin to Android for phones.