Google DeepMind released Gemini Robotics 2 yesterday. It's a three-model family that can make a humanoid robot walk, crouch, and pick things up with some semblance of planning. The release announcement is careful not to oversell it. Good. The honest reporting is the story.
The system can control a humanoid from its feet to its fingertips, coordinate several robots at once, and adapt to a new machine in a few hours. Most robots until now have been either pre-programmed or remote-controlled for narrow tasks. They do not adapt to unpredictable environments, and skills rarely transfer between robot bodies. Gemini Robotics 2 addresses all three at once.
What's odd is what works and what doesn't. Tests showed stronger performance with two-finger grippers than multifinger hands, while Apollo 2 demonstrated full-body tasks including walking, bending and placing objects on shelves. Fine manipulation with all five fingers is the weak spot. Multi-finger dexterity ranges from 32% to 92%. That gap, whether you're trying to pick up a fork or fold a towel, is the actual frontier.
The model architecture is split two ways. Three models ship together: a VLA, an embodied reasoning VLM, and an on-device VLA. The embodied reasoning model handles multi-step planning (what do I do next, and how does the current state change my options). The VLA does the actual motion. Gemini Robotics On-Device 2 can adapt to new two-arm robot designs with fewer than 200 examples. That last number matters, it means you don't need to retrain from scratch when you change hands or bodies.
The access tiers tell a different story. Gemini Robotics ER 2 is public preview; the VLA and on-device models stay gated. Google is opening the planning layer to researchers but keeping the motion controllers behind gates. That's a smart way to gather feedback on the high-level reasoning without releasing hardware control systems into the wild before they're confident.
This is also where Google makes a different bet than you'd see from a pure robotics company. They're not shipping a complete robot. They're shipping intelligence to run on robots that someone else built. One checkpoint drives Apollo 2 with two different hands plus a Franka Duo gripper. That's a real test of generalization, same brain, different bodies, different end-effectors. The fact that it works with some fidelity is the advance.
The weak dexterity numbers are interesting because they're honest. DeepMind could have cherry-picked the two-finger tasks or just not published the multi-finger scores. Instead they showed both. It gives you real information about what's solved and what isn't. A two-finger gripper is fine for many warehouse tasks. Five fingers is what you need for anything humans actually do, sewing, assembling, peeling, the kind of dexterity that takes years to learn. The gap from 92% down to 32% on some tasks is the gap between useful and not yet.
Top comments (0)