Thanks for checking back—I'm here now. Let me give you a proper read on this one. The source describes three models: Gemini Robotics 2 (a vision-language-action model for whole-body and dexterous control), Gemini Robotics ER 2 (an embodied reasoning agent for planning and multi-robot coordination), and Gemini Robotics On-Device 2 (a lightweight version for local inference). The blog post demonstrates whole-body humanoid control—a robot walking to retrieve and place objects—plus dexterous manipulation with five-fingered hands and grippers. They claim adaptation to new robot bodies in "a few hours" with under 200 examples. On safety, DeepMind introduced ASIMOV-Agentic, a benchmark measuring whether the reasoning agent refuses unsafe VLA calls and requests human intervention under uncertainty. The post flags that multi-finger dexterity "remains challenging" and that movement speed needs improvement. What's notably absent: actual success rates for the whole-body tasks (the bar chart references them but the source doesn't include numbers), specifics on what "a few hours" means in wall-clock time, and any discussion of failure modes beyond dexterity difficulty. The source also doesn't detail whether these models can handle novel objects or just trained categories. One sharp question: the source claims the On-Device model "adapts...with just a few hours of adaptation time" but doesn't clarify whether that's supervised fine-tuning or zero-shot. That distinction matters a lot for real deployment.