Gemini Robotics 2 shifts the debate from robot hands to whole-body control

Gemini Robotics 2 is less about a robot answering prompts and more about how far a learned control stack can reach through the body. Google DeepMind frames the release as three connected models: a vision-language-action model for motor control, an embodied reasoning model for planning and interaction, and an on-device VLA model for lower-latency deployment on robots.

The headline capability is whole-body humanoid control. DeepMind says the new system can drive platforms such as Apptronik’s Apollo 2 through tasks that require walking, crouching, stretching, grasping, and placing objects. That matters because many useful environments are built around human motion, not tabletop manipulation. Moving from upper-body demos to feet-to-fingertips coordination is the practical step the post wants readers to notice.

The second thread is dexterity. The company describes Apollo 2 using a five-fingered, 22 degree-of-freedom hand for actions such as tying knots and sealing a ziplock bag, while a Franka Duo setup uses two-finger grippers for tighter packing tasks. Gemini Robotics ER 2 sits above that control layer, interpreting human instructions, planning multi-step tasks, tracking progress, and coordinating multiple robots when one body is not enough.

HN discussion added useful skepticism. Some readers saw the announcement as evidence that Google is advancing on many AI fronts at once: frontier models, smaller fast models, open weights, media generation, and robotics. Others focused on the control stack, arguing that running large language models close to actuation may be too slow or too unstable for practical robotics, with classical control still needed for motion.

Safety is a major part of the release. DeepMind introduced ASIMOV-Agentic, a benchmark for agentic safety orchestration and uncertainty handling, including refusing unsafe tool calls and asking for human intervention when a task may not be feasible. The company also says ER 2 improves human-proximity behavior and safety-stop handling. The result is not a consumer robot launch; it is a clearer view of where embodied AI is trying to merge planning, dexterity, transfer, and safety. Sources: DeepMind and HN discussion.