Gemini 4 Hydrogen Scores 58 on 6-Axis Robot Arm Test
Google says Gemini 4 Hydrogen scored 58 on a real 6-axis robot arm benchmark and launched an Antigravity plugin for hardware setup, simulation, perception and control.

Why it matters
- November 12: Google launched Gemini 4 Hydrogen for fast, efficient agentic workflows.
- Hydrogen scored 58 on Artificial Analysis Robotics 0.1 while controlling a real 6-axis arm through eight tabletop tasks.
- Google said Hydrogen achieved a result similar to Argon at a lower cost per token.
- The Robotics plugin packages four workflows in Google Antigravity: robot bring-up, MuJoCo simulation, perception and control, and application building.
- On ERQA, Hydrogen outperformed Gemini Robotics ER 2 on multi-view correspondence and 3D trajectory reasoning.
By Rebecca Stone
Google reported that Gemini 4 Hydrogen scored 58 on the Artificial Analysis Robotics 0.1 benchmark, controlling a real 6-axis arm across eight tabletop tasks. Google described the result as similar to Gemini 4 Argon’s, but at a lower cost per token.
The benchmark gives Google’s general-purpose robotics claim a physical test. The eight tasks included pouring a drink and packing blocks into a box, according to Google’s developer blog. Hydrogen also retained the arm-control behavior Google previously showed with Gemini 4 Argon and Carbon.
What can Gemini 4 Hydrogen control?
Google says developers can give Hydrogen a goal in plain language and provide a camera view. The model then plans the task, writes and runs control code, and adjusts until the job is done.
“Give it a goal in plain language and a camera view, and it plans the task, writes and runs the control code, and keeps adjusting until the job is done.”
The workflow combines perception, planning and action in one general model rather than requiring Google to present the result as a narrow controller for a fixed set of motions. Google said the Robotics 0.1 result came from a real 6-axis arm, not a simulation.
That claim matters for robotics developers because Google’s release addresses both the model and the machinery around it. The company paired Hydrogen’s control loop with a plugin that installs hardware drivers, builds simulations, generates application code and closes the loop between perception and physical motion.
Where does Hydrogen fit among Google’s models?
Google launched Gemini 4 Hydrogen on November 12 as a general-purpose, readily available model designed for speed and efficiency in agentic workflows. Its initial release targeted everyday execution work, including writing code, running tests and verifying changes.
Google said teams still use frontier models such as Carbon and Argon for architectural judgment when they design systems or untangle subtle bugs. Hydrogen fills a different role: fast, disciplined execution.
“Hydrogen is built to stand on its own as a model for everyday work — or to pair with Carbon and Argon as the high-efficiency execution engine in multi-model workflows.”
Google also brought capabilities from Gemini Robotics ER 2, its Embodied Reasoning model family, into the core Gemini 4 series. The company said Hydrogen beats Gemini Robotics ER 2 on nearly every axis and recommends it to developers now using ER 2.
On the public ERQA benchmark, Google said Hydrogen outperformed Gemini Robotics ER 2 on spatial-reasoning work, including multi-view correspondence and 3D trajectory reasoning. Those are the two named areas in Google’s comparison; the company did not publish a task-by-task breakdown in the cited post.
Google’s framing presents embodied reasoning as a capability of the wider Gemini 4 line rather than a product reserved only for a dedicated robotics model. It also places Hydrogen against Google’s own earlier robotics work, making the release a model-consolidation claim as much as a new-model launch.
What does the Antigravity Robotics plugin do?
Google launched the Robotics plugin for Google Antigravity alongside its robotics push. Google described the plugin as a set of skills for hardware onboarding, simulation and robot control that turns Antigravity into a one-stop robotics workbench.
Google said a developer can move from a robot in a box to a working application in an afternoon. Installation takes place inside Antigravity: the developer opens Customizations and searches for Robotics on the Build With Google shelf.
The plugin addresses a stack that Google said robot applications have long required developers to assemble manually. That stack includes drivers, calibration scripts, simulators, perception code and a user interface. In Google’s account, the work demands robotics expertise and substantial time before an application can run on hardware.
The plugin packages four pieces of that process as skills that Gemini can use on the developer’s behalf:
- Robot bring-up: Installs drivers, discovers cameras, calibrates the arm and verifies safe motion, beginning from a USB cable.
- MuJoCo in Antigravity: Builds and previews simulated scenes and assets inside Antigravity so developers can test ideas before touching hardware.
- Perception and control: Uses Gemini’s spatial understanding to find objects, plan grasps and control the robot in a closed loop.
- Application building: Uses Gemini’s coding skills to generate the controller and simulation and complete a full Real-to-Sim-to-Real application workflow.
Those pieces map to a concrete development sequence. A developer starts with the physical arm and its cameras, prepares a simulated version, gives the system a task, and then returns the generated controller to real hardware. Google said Antigravity handles each stage through the same interface.
That sequence matters because Google’s plugin addresses setup steps beyond model inference. A working system also needs recognized devices, completed calibration, checked motion and a controller tied to physical hardware. Google placed those operations inside the same environment that writes the controller.
Why does Google’s general-purpose robotics push matter?
WebProNews connected Google’s launch with OpenAI’s decision to disband its dedicated robotics team. The publisher reported that many industry observers viewed OpenAI’s move as recognition that general-purpose AI models had become so adept at working with robotics that dedicated models were no longer needed to the degree once thought.
The comparison frames Google’s release against a broader shift in AI development. Instead of treating robotics as a separate model category, Google is positioning Hydrogen as an everyday model that can control hardware, work with embodied-reasoning models and sit inside a general software-development environment.
OpenAI’s organizational decision does not prove that specialist robotics systems will disappear. The source offers no adoption data, deployment totals or customer results for either company. It does show Google making the opposite organizational bet in public: distribute general-purpose robotics capabilities through a fast model and package the supporting tools for immediate use.
For the robotics market, Google’s release puts the development workflow alongside model performance. Google says its plugin can shorten the path from an unconfigured arm to an application, while Hydrogen handles the language-to-control loop. The package now competes on setup time and execution efficiency, not only benchmark scores.
Google says building robot applications has required substantial expertise and time. Its plugin is designed to reduce that burden by bringing hardware configuration, simulation, perception, control and application code into one environment. Hydrogen supplies the general model that can work across those stages.
What should developers test next?
Google has supplied three measurable starting points: a score of 58 on Artificial Analysis Robotics 0.1, physical control of a 6-axis arm across eight tabletop tasks, and an ERQA result above Gemini Robotics ER 2 on multi-view correspondence and 3D trajectory reasoning. The Robotics plugin adds a test of whether its onboarding and simulation skills work with hardware beyond the cited demonstration.
The next release of evidence should focus on those practical points. Developers need to know whether Hydrogen maintains its score as tasks and hardware change, whether the plugin removes setup work reliably, and whether lower per-token cost persists under repeated perception-and-control cycles. Those tests will determine whether Google’s unified robotics approach can move beyond the eight-task demonstration.
Source: webpronews.com
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
728 articles
Related articles
- Google's Gemini 4 Argon Wants to Be the One Model Robots Run On
- Google DeepMind's Gemini Robotics 2.5 hits 96% in Mercedes factory pilot
- Gemini Robotics-ER 1.6 adds instrument reading for Boston Dynamics' Spot
- Google Launches Gemini Robotics ER 2 With Multi-Robot Collaboration
- Google Launches Gemini 4 Hydrogen, Beating Argon on Intelligence Index