Google's Gemini 4 Hydrogen Turns Robot Arm Setup From Weeks to Hours
Google's fast Gemini 4 Hydrogen model now comes with an open-source Antigravity Robotics plugin. It handles robot setup, MuJoCo simulation and camera-guided control from plain-language prompts.
- Google launched an open-source Robotics plugin for Antigravity built around the fast Gemini 4 Hydrogen model.
- Hydrogen scores 58 on AA-Robotics v0.1, matching Gemini 4 Argon at lower cost per token.
- Hydrogen beats specialist Gemini Robotics ER 2 on ERQA spatial reasoning, averaging 84.8 versus 75.5.
- The plugin handles driver install, camera discovery, calibration, MuJoCo simulation and closed-loop arm control from chat prompts.
- Demos: YAM arm calibrated in about 10 minutes, conveyor snack sorter built in about an hour.
- Hydrogen costs $1.50/$7.50 per million input/output tokens. Antigravity access requires Google AI Pro or Ultra.
Google has released a Robotics plugin for Antigravity that uses Gemini 4 Hydrogen to configure supported robot arms, build MuJoCo simulations, and run camera-guided control from chat. Google says the tooling can compress weeks of driver setup, calibration, simulation, and perception work into an afternoon.
The workflow starts with an arm and depth camera connected to a laptop. A developer describes the task, and the agent brings up the hardware, creates a simulation, generates control code, and executes the task. During closed-loop control, the model watches the camera feed and continually adjusts the arm’s motion based on what it sees.
Hydrogen absorbs the robotics specialist
Google designed Hydrogen as a fast, lower-cost model for agentic workflows, while the larger Gemini 4 Argon and Carbon models handle more demanding reasoning. API pricing remains constant across its 1 million-token context window at $1.50 per 1 million input tokens and $7.50 per 1 million output tokens.
The independent AA-Robotics v0.1 benchmark gives a model control of a real six-axis arm and tests it on eight tabletop tasks, including pressing a button, pouring a drink, and writing "hello world" with a marker. Google reports that Hydrogen scored 58, roughly matching Gemini 4 Argon while using less expensive tokens.
Google previously offered object localization, grasp-point selection, and three-dimensional motion reasoning through Gemini Robotics ER 2. Hydrogen incorporates those capabilities into the general Gemini 4 line. On the public ERQA benchmark, Google reports an average score of 84.8 across 14 embodied-reasoning categories, compared with 75.5 for ER 2. The tasks include matching information across camera views and reasoning about three-dimensional trajectories. Google recommends that ER 2 users evaluate Hydrogen for the same workloads.
Four skill groups handle the setup
The plugin packages instructions and tools into reusable “skills” that Antigravity invokes during each stage of a robotics project:
- Robot bring-up: Installs drivers, discovers cameras, calibrates the arm, and runs basic motion checks from a USB connection.
- MuJoCo simulation: Runs MuJoCo beside the application code, with MuJoCo Studio and MuJoCo Playground included.
- Perception and control: Uses Hydrogen’s spatial reasoning to locate objects, choose grasp points, plan movement, and correct the arm’s actions from camera feedback.
- Application generation: Produces the controller, simulation, and supporting application code. Developers can reconstruct a physical workspace in simulation, test the behavior, and deploy it to the arm.
Out-of-box support covers three robot configurations:
| Hardware | Positioning |
|---|---|
| I2RT YAM | Arm used in Google’s bring-up and sorting demonstrations |
| SO-ARM101 | Lower-cost arm from the LeRobot ecosystem |
| Franka FR3 Duo | Industrial dual-arm configuration used in the chess demonstration |
The plugin is available on GitHub under the Apache 2.0 license. Developers can write skills for additional hardware and share them through the project’s GitHub discussions. Hardware outside the supported list will still require integration work, testing, and device-specific safety controls.
Prompts build the test cell
Google’s I2RT YAM demonstration completed software installation, camera discovery, arm calibration, and test movements in roughly 10 minutes. Once configured, the system handled prompts such as “pick up a pen” and “stack snacks” within minutes.
The plugin also supports “vibe sim,” Google’s name for generating simulation assets from a written description, workspace photograph, or product-page link. Developers can adjust geometry, object positions, and physical properties through chat before transferring a tested behavior to the physical workspace.
Google presents three demonstrations of the complete workflow:
- Snack-sorting conveyor: Hydrogen converts a whiteboard sketch and scene photographs into a MuJoCo simulation, evaluates grasps across repeated runs, and deploys the controller to a real arm. The developer then adjusts gripper force and sorting rules in plain language. Google reports an end-to-end setup time of about one hour.
- Self-fulfilling soap store: Hydrogen builds a web storefront and connects it to two YAM Pro arms. Incoming orders trigger the arms to select soap bars, pack them in a bin, and add a printed thank-you note.
- Chess robot: The FR3 Duo creates a simulated board and then plays against itself or a person.
The soap-store demonstration illustrates Google’s cost argument. A physical workflow can generate a long sequence of perception, planning, and correction calls, so Hydrogen’s lower token price reduces the expense of operating the agent throughout the day.
Three routes into the stack
Developers can access Hydrogen for robotics through Antigravity, the Gemini API, or AI Studio:
| Environment | How to start |
|---|---|
| Antigravity | Select gemini-4-hydrogen, open Customizations, and find Robotics on the Build With Google shelf. Access requires a paid Google AI Pro or Ultra subscription. |
| Gemini API | Use gemini-4-hydrogen through the Interactions API or the google-genai SDKs. |
| AI Studio | Test spatial prompts, live camera streams, and structured tool calls in a browser. |
A basic SDK request can ask Hydrogen to identify graspable objects and return normalized image coordinates. The exact response schema should be checked against the API documentation.
from google import genai
from google.genai import types
client = genai.Client()
with open("workspace.jpg", "rb") as image_file:
image = image_file.read()
response = client.models.generate_content(
model="gemini-4-hydrogen",
contents=[
types.Part.from_bytes(
data=image,
mime_type="image/jpeg",
),
(
"Point to every graspable snack. Return JSON as "
'[{"point": [y, x], "label": "string"}]. '
"Normalize coordinates from 0 to 1000."
),
],
config=types.GenerateContentConfig(
response_mime_type="application/json",
),
)
print(response.text)The benchmark exposes a production gap
The independent AA-Robotics results show how far model-controlled arms remain from routine production use. Hydrogen scored 58, while the leaderboard’s top model, GPT-6 Astra, scored 69. The benchmark authors reported latency and cost nearly two orders of magnitude worse than human baselines. Models averaged about 30 minutes per task, compared with 20 seconds for a person, and Argon took 52 minutes per task.
Argon retained its target object in only 36% of grasp attempts. Recorded failures included closing the gripper on empty space, knocking over object stacks, and declaring tasks complete prematurely. Hydrogen’s similar aggregate score does not establish an equivalent grasp-success rate, since Google has not published that measurement. Developers will need to test repeated grasps, recovery behavior, and completion detection with their own objects and camera positions.
Google supplied the reported 10-minute bring-up, one-hour sorting cell, and ERQA results. These figures have not been independently reproduced across varied lighting, clutter, hardware wear, or unfamiliar objects. The plugin’s motion checks also do not replace safety-rated controllers, physical stops, guarded workspaces, or application-specific risk assessment.
Where the plugin fits today
The combined release shows three parts of Google’s robotics strategy:
- General models are absorbing specialist skills: Hydrogen brings the spatial capabilities of Robotics ER 2 into the main Gemini model family and reports higher ERQA scores.
- Coding agents can orchestrate physical systems: The model generates and executes control code, observes the result through cameras, and revises the next action. This architecture can support reactive control without requiring a separately trained end-to-end policy for every task.
- Simulation is becoming conversational: Prompt-generated MuJoCo scenes reduce the manual work required to create robot descriptions, place assets, and tune physical properties.
Current evidence supports the plugin as a prototyping tool for pick-and-place cells, laboratory automation, teaching, and research arms such as the SO-ARM101. Production lines impose stricter requirements for cycle time, deterministic behavior, grasp reliability, fault recovery, and safety certification.
A practical evaluation should measure repeated grasp success, end-to-end latency, token cost, recovery from misplaced objects, simulation-to-reality error, and behavior at workspace boundaries. Those results will determine whether Antigravity saves setup time on a specific robot without shifting the work into debugging generated code and failed motions.