What Gemini Robotics 2 actually is: three models doing three jobs

Google DeepMind's Gemini Robotics 2 is not a robot. It is three software models, released together in a single company blog post. One turns what a robot sees and what it is told into motor commands. One breaks a multi-step task into steps. One runs locally on the robot without a cloud round-trip. The point of releasing them as a stack is to split the parts of "intelligence" that robotics has historically lumped together.

A vision-language-action model (VLA) is the part of the stack that looks at a camera feed and the user's instruction, then outputs motor commands: joint angles, gripper pressure, base velocity. A vision-language model (VLM) is what most people have used through chatbots: it sees an image and answers in text. An embodied-reasoning (ER) model is a VLM tuned to act as a planner, decomposing "tidy the room" into a sequence of graspable subtasks. Gemini Robotics 2 is the VLA. Gemini Robotics ER 2 is the planner. Gemini Robotics On-Device 2 is a lighter VLA that runs on the robot itself rather than in a data center.

What changes with "whole-body control" is the addressable surface. Earlier manipulation systems usually drove a gripper on a fixed arm. DeepMind says its stack can drive a humanoid from feet to fingertips, which in practice means the robot can crouch to pick something off the floor, stretch to reach a high shelf, and coordinate two hands on a deformable object like a towel. The vendor post frames this as a single intelligence layer rather than separate locomotion and manipulation controllers, and pairs it with a claim that the VLA can adapt to an entirely new robot body in "a few hours" of training. That number, per the accompanying preprint, is vendor-stated and not independently benchmarked.

A model that runs locally on the robot does not need a network connection to function, and the latency between an action and a result is the difference between a robot that catches a falling object and one that drops it. DeepMind is also claiming multi-robot team collaboration, with separate robots coordinating on a shared task, though the company has not published the failure modes or success rates that would let an outsider tell whether the teams are doing genuine coordination or are being orchestrated by a single planner.

Google DeepMind has shipped a video and a vendor blog post, not peer benchmarks against competitors like 1X, Figure, Physical Intelligence, NVIDIA GR00T, or the Tesla Optimus stack, and has not disclosed failure modes or named any production partner. Hacker News commenters are split in the way the field usually splits: impressed by the dexterity footage, skeptical that tidy-room demos will hold up in unstructured homes. Both reactions are reasonable, and the source itself does not resolve the gap.

What to watch next: a third-party benchmark of adaptation time and dexterity, and any production deployment from a named partner other than DeepMind's own demo rooms. Until then, the test for any future Gemini Robotics post is the same as the test for this one: does it show what the model fails at, not just what it succeeds at?