A light bulb unscrewed itself with 92 percent reliability on July 30. The hand doing the work belonged to Apptronik's Apollo 2 humanoid, controlled by a new AI system from Google DeepMind called Gemini Robotics 2. The task sounds trivial until you consider what it requires: the robot's feet must balance while its torso rotates; its shoulder and elbow must position the hand; its wrist must twist; five fingers must grip without crushing the bulb; and the entire chain must coordinate in real time without a human in the loop. Previous versions of Gemini Robotics controlled only the upper body for tabletop work. This one controls the whole skeleton.
Google DeepMind released three models in the Gemini Robotics 2 family on July 30 and 31, 2026. Gemini Robotics ER 2 runs in the cloud and is available now through Google AI Studio and the Gemini Enterprise Agent Platform. Gemini Robotics 2 and Gemini Robotics On-Device 2 went into private preview with more than 100 early-access partners: Apptronik, Boston Dynamics, and Agile Robots among them. The on-device version is the one that matters most for commercialization. It runs directly on a robot's onboard computer, requires no cloud connection, and can be adapted to a new robot design with fewer than 200 training examples. That last number is the real story. Until now, teaching a custom robot to do a new task meant collecting thousands of hours of demonstration data or spending months on reinforcement learning. Gemini Robotics On-Device 2 collapses that timeline to hours.
The performance numbers reveal what becomes possible when you move from upper-body control to whole-body coordination. On Apollo 2, the system achieved a 68.4 percent success rate picking objects from a table, 45.7 percent from the floor, and 76.3 percent from a shelf. Those are not perfect, but they represent the first time a generalist robot AI has handled the balance and limb coordination required for floor-to-shelf manipulation, the kind of work that robots are actually deployed to do in warehouses and factories. The system also demonstrated five-finger dexterity on SharpaWave's 22-degree-of-freedom hand, tying knots and sealing ziplock bags. That combination of strength, precision, and adaptability has been the holy grail of robotics AI for three years. Google just shipped it.
Carolina Parada, head of robotics at Google DeepMind, framed the move plainly: 'Our goal is to bring AI into the physical world and then build the intelligence layer that can be used by every robot.' That is not a technology pitch. That is a platform play. Google is not trying to dominate robot hardware. It is trying to own the software stack that every robot builder will eventually license. The signal is the timing and the partnership composition. Apptronik, Boston Dynamics, and Agile Robots are not competitors in the humanoid market; they are the three leading commercial platforms. Google is not fragmenting the market by backing one. It is standardizing it by offering all of them the same capability. Whichever hardware maker does not integrate Gemini Robotics 2 within six months faces a problem: customers will compare robots side by side on the same AI stack and make hardware choices based on physical design, price, and reliability, not on whose proprietary software stack is better. That conversation is already over.
Google also released ASIMOV-Agentic, a safety benchmark for agentic robots that measures a system's ability to refuse unsafe actions, recognize when a task cannot be completed safely, and request human intervention when appropriate. This is not a regulatory requirement. It is a cultural reset. Every robot sold through one of these three partnerships will have the same safety model, the same refusal mechanisms, the same request-for-human-intervention logic. In effect, Google just standardized robot safety across the entire commercial humanoid market. No startup or alternative platform can ship without addressing why their safety model is different.
Watch three things to see whether this becomes the de facto standard. First: within 90 days, count how many new robot hardware announcements explicitly claim compatibility with Gemini Robotics 2. Second: track whether any major industrial customer (automotive, logistics, semiconductor) makes a public purchasing decision conditional on Gemini Robotics 2 compatibility. Third: monitor whether any rival AI lab, Elon Musk's xAI, OpenAI's robotics push, or an established automation player, announces a competing whole-body control stack with on-device adaptation in the next six months. If none of these signals appear, Google has won the platform layer, and every robot company has just become a hardware manufacturer selling into a Google-controlled software stack.