Gemini 4 Hydrogen: a faster model for general purpose robotics
Google launched Gemini 4 Hydrogen alongside an Antigravity robotics plugin that sets up physical robot arms and runs MuJoCo simulations in ten minutes.
By Dillip Chowdary • Nov 18, 2026 • Source: Google Developers Blog

At a glance
- Product
- Gemini 4 Hydrogen
- Developer
- Benchmark score
- 58 on Artificial Analysis Robotics 0.1
- Setup time
- Approximately 10 minutes for YAM arm
- Availability
- Google AI Studio and Interactions API
- Platform requirements
- Google AI Pro or Ultra for Antigravity
In this post
Google launched Gemini 4 Hydrogen, introducing an embodied reasoning model designed for general-purpose robotic control alongside a dedicated Robotics plugin for Google Antigravity. As detailed in Google Developers Blog's report, the release pairs the faster, lower-cost model with an integrated development environment to automate hardware configuration, physics simulation, and closed-loop execution. The software package targets roboticists and software engineers who need to assemble physical automation tasks without handcrafting custom device drivers or separate control pipelines.
This report examines the architecture of Gemini 4 Hydrogen, its benchmark performance across physical manipulation suites, and its integration inside Google Antigravity. It breaks down the system onboarding flow, MuJoCo physics generation, and two real-world implementations involving dual-arm retail fulfillment and automated conveyor sorting.
01Benchmark scoring and physical reasoning
Gemini 4 Hydrogen integrates embodied reasoning capabilities directly into the core Gemini 4 model family. On the public ERQA benchmark, the model outperforms Google's earlier Gemini Robotics ER 2 system on spatial reasoning evaluations, including multi-view correspondence and 3D trajectory tracking. It interprets plain-language instructions alongside live camera feeds to generate machine control code and maintain task loops.
On the Artificial Analysis Robotics 0.1 benchmark, the model achieved an overall score of 58 across eight tabletop manipulation tasks. This test suite evaluates physical arm operations, including pouring liquids and packing solid blocks into containers. The resulting performance matches the larger Gemini 4 Argon model while running at a reduced cost per token.
| Model | Evaluation Suite | Task Scope | Benchmark Score |
|---|---|---|---|
| Gemini 4 Hydrogen | Artificial Analysis Robotics 0.1 | 8 tabletop manipulation tasks | 58 |
| Gemini 4 Argon | Artificial Analysis Robotics 0.1 | 8 tabletop manipulation tasks | 58 |
| Gemini Robotics ER 2 | ERQA Benchmark | Multi-view correspondence, 3D trajectories | Lower than Hydrogen |
02Antigravity robotics plugin architecture
Building robotic applications traditionally requires engineers to manually link hardware drivers, calibration routines, physics simulators, and vision models. To eliminate this fragmented setup, Google launched the Robotics plugin on the Build With Google shelf inside Google Antigravity. The plugin provides automated skills that handle hardware onboarding over standard interfaces, scene building, and continuous motion tracking.
- Hardware discovery connects over USB to identify external depth sensors and install device drivers.
- Kinematic calibration executes automated motor sweeps to verify safe joint limits and motion envelopes.
- Simulation asset generation loads workspace photos or documentation into MuJoCo physics scenes.
- Closed loop control issues trajectory adjustments and joint torque parameters during task runs.
The integrated MuJoCo engine allows developers to construct, inspect, and audit physical environments alongside their application code. Users can describe items in natural language, import specification documents, or upload reference images to construct simulation assets. Physical characteristics, object geometries, and target positions can then be adjusted directly inside the environment before sending motion plans to physical hardware.
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
03Hardware bring-up and calibration workflows
Hardware onboarding through Antigravity reduces physical configuration times down to a few terminal prompts. In a demonstration using an I2RT YAM arm and an external depth camera connected via USB, the system scanned the bus, configured drivers, calibrated joint axes, and executed test sweeps in approximately 10 minutes. Similar bring-up sequences were demonstrated on the low-cost SO-ARM101 arm and the dual-arm industrial FR3 Duo.
Once basic kinematic checks pass, operators can issue plain-language commands to command preliminary arm routines. These initial scripts verify that motor controllers and camera hand-eye transformations align properly within the workspace coordinate frame. Standard verification actions include commanding greeting gestures, picking up table stationery, or stacking small objects.
Note: Access to Google Antigravity and the Robotics plugin requires an active paid Google AI Pro or Ultra subscription.
04Conveyor sorting and store fulfillment
To demonstrate practical deployment speeds, Google detailed an automated conveyor-sorting cell constructed from a whiteboard drawing. A developer provided an overhead photograph of a conveyor line carrying mixed snack items, prompting Gemini 4 Hydrogen to build a representative MuJoCo simulation. The system iteratively evaluated grasp vectors in software before exporting the control policy to physical arm hardware.
The developer adjusted parameters such as gripper closure pressure and item sorting criteria using conversational prompts. The complete pipeline transitioned from a whiteboard sketch to a production cell sorting snacks off a live belt within an hour. Hydrogen continuously monitored camera frames to recalculate pickup timing as items advanced down the belt line.
A separate implementation deployed two YAM Pro arms to operate an automated order-fulfillment storefront for a specialty soap business. Gemini 4 Hydrogen generated the web storefront interface, mapped incoming customer orders to inventory bins, and orchestrated the dual-arm picking system. Upon receiving an order, the arms positioned packing boxes, retrieved the specified soap bars, and packed printed receipts.
05SDK access and platform availability
Gemini 4 Hydrogen is accessible through the Interactions API and is supported on launch day across the official google-genai software development kits. Developers can experiment with spatial reasoning prompts, video camera inputs, and structured function calling inside Google AI Studio. System prompts can reference model identifier gemini-4-hydrogen directly.
Inside Google Antigravity, developers can select the model from the model selector and install the Robotics plugin from the Customizations menu. The tooling supports complex game logic alongside industrial tasks, such as commanding an FR3 Duo arm to construct a simulated chessboard and compete against human players or self-play instances. Community discussion boards and custom skill contributions are hosted on the plugin repository on GitHub.
Frequently asked questions
What hardware platforms support the Antigravity Robotics plugin?
The plugin has verified support for the I2RT YAM arm, the YAM Pro dual-arm setup, the low-cost SO-ARM101, and the industrial FR3 Duo.
How does Gemini 4 Hydrogen compare to Gemini 4 Argon?
Hydrogen matches Argon with a score of 58 on the Artificial Analysis Robotics 0.1 benchmark while offering lower token inference costs and faster response times.
Which subscription tier is needed to use the Robotics workbench in Antigravity?
Developers must have an active paid Google AI Pro or Ultra subscription to access the Antigravity workspace and its Robotics plugin.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement