Google DeepMind has released Gemini Robotics 2, a vision-language-action model that gives humanoid robots full-body control. The system enables robots to walk, crouch, stretch, balance and manipulate objects while completing multi-step tasks. In demonstrations, the model directed Apptronik's Apollo 2 humanoid to pick up a watering can, walk across a room and place it on a lower shelf. The same checkpoint also controlled other platforms such as Franka Duo. Dexterity improvements allow five-fingered hands to tie trash bags, seal Ziploc bags and unscrew light bulbs. Two companion models accompany the release: Gemini Robotics ER 2 for high-level reasoning and long-horizon planning, and Gemini Robotics On-Device 2, which runs locally and can adapt to new dual-arm designs with fewer than 200 training examples collected in a few hours. Google introduced the ASIMOV-Agentic benchmark to evaluate safety behaviours such as refusing unsafe actions and requesting human help.
Google DeepMind has released Gemini Robotics 2, a vision-language-action model that gives humanoid robots full-body control. The system enables robots to walk, crouch, stretch, balance and manipulate objects while completing multi-step tasks. In demonstrations, the model directed Apptronik's Apollo 2 humanoid to pick up a watering can, walk across a room and place it on a lower shelf. The same checkpoint also controlled other platforms such as Franka Duo. Dexterity improvements allow five-fingered hands to tie trash bags, seal Ziploc bags and unscrew light bulbs. Two companion models accompany the release: Gemini Robotics ER 2 for high-level reasoning and long-horizon planning, and Gemini Robotics On-Device 2, which runs locally and can adapt to new dual-arm designs with fewer than 200 training examples collected in a few hours. Google introduced the ASIMOV-Agentic benchmark to evaluate safety behaviours such as refusing unsafe actions and requesting human help.