AGENTIC COMMONSAI industry briefings

繁中EN

TagGeminiPublished

All English articles

Gemini 4 Hydrogen: fast robotics control in Antigravity

Summary: Google announced Gemini 4 Hydrogen on November 17, 2026: a faster, cheaper member of the Gemini 4 family that carries the embodied-reasoning skills of its robotics models, plus a new Robotics plugin for Google Antigravity that turns the IDE into a robotics workbench. The claim worth your attention is the loop length — robot in a box to a working application in an afternoon — and the cost of finding out whether that’s true for you is one paid subscription and a cheap arm.

What Hydrogen actually changes

Until now, the embodied-reasoning capabilities lived in Google’s specialized robotics models — the line that grew out of what we covered when Gemini Robotics 2 gave robots whole-body intelligence. Hydrogen folds those skills into the core Gemini 4 series, so a plain-language goal plus a camera view is enough: the model plans the task, writes and runs the control code, and keeps adjusting until the job is done.

The headline number is vendor-measured, so read it that way. On the Artificial Analysis Robotics 0.1 benchmark, Google reports Hydrogen scores 58 controlling a real 6-axis arm through eight tabletop tasks, similar to Gemini 4 Argon at a lower cost per token. On the public ERQA benchmark, Google says Hydrogen beats its own Gemini Robotics ER 2 on spatial reasoning such as multi-view correspondence and 3D trajectory reasoning — which is why Google recommends ER 2 users try Hydrogen for those workloads. Every one of these numbers is Google’s own; no independent measurement appears in the announcement.

The plugin kills the glue work

The model is half the story. Building a robot application has traditionally meant stitching together drivers, calibration scripts, simulators, perception code, and a UI — work that needed robotics expertise. The new Robotics plugin packages that into skills Antigravity runs on your behalf:

  • Robot bring-up installs drivers, finds cameras, calibrates the arm, and verifies safe motion from a USB cable alone.
  • MuJoCo in Antigravity builds and previews simulated scenes inside the IDE, so you can validate a policy in sim before touching hardware.
  • Perception and control uses Gemini’s spatial understanding to find objects, plan grasps, and close the control loop.
  • Application building generates the controller and simulation, completing the real-to-sim-to-real path in one environment.

The trade-off is that this workbench lives inside Antigravity, which requires a paid Google AI Pro or Ultra subscription — the workflow isn’t runnable on the free tier.

How fast the demo loop actually runs

The announcement walks through the full path. A developer plugs an I2RT YAM arm and an external depth camera into a laptop and asks Hydrogen to bring it up; the plugin installs software, finds the camera, calibrates, and runs test motions, with the arm ready for tasks in about 10 minutes. The same flow brings up a low-cost SO-ARM101 and an industrial FR3 Duo.

The flagship example is a snack-sorting conveyor: a whiteboard sketch becomes a MuJoCo simulation built from one prompt and scene images, the grasping policy is validated across multiple sim runs, then deployed to the real arm — and nuances like gripper closure force get tuned by chat. Total time, end to end: about an hour. Two longer-running demos show where Hydrogen’s efficiency is meant to matter: a soap storefront connected to two YAM Pro arms that fulfills orders all day, and an FR3 Duo that simulates a chessboard and then plays.

Our take: act now, as an experiment — not as a commitment. If you’re a small team prototyping physical automation and you already own or can cheaply buy a compatible arm, this is exactly the cheap, reversible test worth running early: one paid subscription afternoon tells you whether the 10-minute bring-up and one-hour sorter hold on your hardware. But treat every number as Google measuring its own demo on three specific arms (YAM, SO-ARM101, FR3 Duo), and name your assumption: if your task isn’t tabletop manipulation in the benchmark’s style, the 58 score tells you little. We’d skip this if your product ships no hardware, or if you already run production robots on your own control stack and evaluation. What would change our position: independent results on other arms, or Google publishing usage-based pricing and availability for running these applications beyond a demo.

Getting started, and what’s still open

Getting started is deliberately short. In Antigravity, select gemini-4-hydrogen, open Customizations, and search for Robotics on the Build With Google shelf; the announcement’s suggested first prompt is “Bring up my SO-101 arm and make it wave.” Hydrogen is also available via the Gemini Interactions API with day-one support in the google-genai SDKs, and in Google AI Studio for spatial prompts, live camera streams, and structured tool calls in the browser.

What the announcement doesn’t state: per-token pricing for Hydrogen, regional availability, or how the system behaves when a real-world deployment drifts from its validated sim. For an afternoon experiment none of that blocks you; for anything a customer depends on, it should.

Sources

Original post
AGENTIC COMMONSOperated by PHLEGON LABS
ShareXEmail
Support us

Related reading

  1. Gemini 4 Neon: pricing, benchmarks, and where it fits

    Gemini 4 Neon ships in the paid Gemini API at $3 per million input tokens. The benchmarks are Google's own: run a test this week, don't migrate yet.

    Gemini

  2. Gemini 4 Carbon ships at Argon pricing: judge it per task

    Gemini 4 Carbon is now open to paying developers at Argon's $2/$10 intro pricing — with a heavier token burn per task.

    Gemini

  3. What Google's Argon Release Actually Changes for the Way You Ship

    Gemini 4 Argon launches at $2/$10 per million tokens with benchmarks, agent speedups, and a clear case for rethinking model selection.

    Gemini