
Google DeepMind has launched Gemini Robotics 2, a vision-language-action model capable of whole-body control for humanoid robots like Apptronik’s Apollo 2.
The model enables robots to perform complex tasks such as tying knots or sealing bags using a 22 degree-of-freedom hand, and supports multi-robot collaboration.
DeepMind also introduced the ASIMOV-Agentic benchmark to measure safety and uncertainty resolution, ensuring robots can request human intervention when tasks are deemed unsafe.