Google DeepMind has released Gemini Robotics 2, a vision-language-action (VLA) model that converts what a robot sees and feels into motor commands for the whole body, from feet to fingers. The model makes a humanoid walk and bend, manipulates objects in cluttered rooms, and coordinates multiple robots working together. It also runs on-device and adapts to new robotic bodies in a few hours.

The announcement comes three days after we reported on SceneActBench, the benchmark that evaluates vision-language models on coordinated actions with multiple objects in 3D scenes. Eleven models tested, scores between 38 and 50 out of 100: none performs well across the board. Now there’s a lab model that attempts to do exactly those things, on a real body.

Why it matters to you. If you work with AI agents, so far you’ve measured success on text and code. VLMs that describe images stay in the same family. The leap toward agents that manipulate the physical world shifts the question: how coordinated is their action in space matters as much as what they know how to say. Benchmarks are catching up. Gemini Robotics 2 shows that labs are investing seriously, with models that control fine-grained movement and multi-robot collaboration.

On the same front, MarbleOS (Show HN, 113 points) proposes a graphical interface for AI agents working with files, tools, and visible outputs instead of buried chat threads. It’s an early demo, but the signal is consistent: when the agent does things in the world instead of writing, the interface needs to show the work in progress.

Limitations remain substantial. Gemini Robotics 2 is a lab announcement, still without a purchasable product. Benchmarks on 3D scenes record low scores for everyone. Independent reproducibility is missing: for now, numbers on adaptation to new bodies come from DeepMind itself.

In detail

Context: for months we’ve discussed agents working on text and code. Papers on generative world models had shown the direction, training robots in simulation rather than the real world. Two frameworks on spatial reasoning had marked the limit: VLMs that describe images don’t reason well about physical space. Gemini Robotics 2 attempts to close that gap with a different architecture.

What changes technically. A vision-language-action model doesn’t stop at description. It takes as input the video stream from the robot’s cameras and natural language instructions, and outputs commands for every motor joint. The model controls the entire body: walking, fine object manipulation, limb coordination. DeepMind claims adaptation to new robotic bodies in a few hours and on-device execution without the cloud.

The novelty versus the first generation of Gemini Robotics is whole-body control. Before, VLAs managed an arm or gripper; now they cover from walking to fine manipulation. It adds multi-robot collaboration: two robots dividing work to tidy a room.

Where sources diverge and what remains open. DeepMind’s post is an announcement with video demos, not a paper with reproducible metrics. Numbers on adaptation to new bodies (a few hours) and on-device execution are laboratory claims, not independently verified. SceneActBench, released on arXiv three days earlier, shows the opposite picture: when you objectively measure how VLMs coordinate actions on 3D scenes, scores range from 38 to 50 out of 100. The gap between the announcement and independent measurement is wide.

MarbleOS, reported on Hacker News the same day, touches a connected but different node: if the agent operates on files, tools, and output instead of only answering in chat, an interface that shows work in progress is needed. It’s a beta with 113 points on HN, not a mature product, but the interface problem for agents acting in the world is real.

What not to conclude. A robot that walks and manipulates objects in a video demo doesn’t mean VLAs are production-ready. Benchmarks say otherwise. The useful signal is different: labs are investing in models that act in the physical world, and tools to measure them arrive in parallel. For anyone building agents today on text and code, it’s worth following this line: the evaluation metrics you’re using now may not suffice when your agent needs to operate in space rather than on the page.