Skip to content

Tools

Google adds Robotics to Antigravity, linking robot-arm setup with simulation

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. The November 17 Robotics plugin brings the existing Hydrogen model into robot workflows.
  2. Builders can work on simulation and physical control code, including simulation-only projects.
  3. Physical motion requires approval and checks; personal Hydrogen access covers paid Pro and Ultra.
PracticalDevelopers
A robot arm holds its gripper above wooden blocks beside a laptop showing a wireframe of the scene
AI-generated conceptual image: a robot arm and a laptop model represent the link between simulation and physical control · AI Ping

Google released a Robotics plugin for Antigravity on November 17, combining robot-arm setup, simulation and physical control in its development tool. Gemini 4 Hydrogen can turn a task goal into control code, test it in simulation and apply the code to a real arm. Builders without hardware can work entirely in simulation.

The release packages robotics workflows around the existing Hydrogen model. The hardware instructions cover I2RT YAM, SO-101 and Franka FR3 Duo arms. For personal accounts, Hydrogen is available on paid, non-trial Google AI Pro subscriptions and on Ultra. The official model table also lists Gemini Enterprise support. The public repository contains task instructions and helper code for hardware setup, perception and motion verification.

In Google's examples, bringing up a YAM arm through calibration and test movements took about ten minutes. A conveyor snack-sorting application went from a sketch through simulation to physical deployment in roughly an hour, the company said. Those timings describe its individual demonstrations.

Checking motion before physical deployment

Simulation includes checking whether objects behave physically. The published audit instructions cover objects floating above or sinking into the floor, collision shapes that fail to match visible geometry, and broken joints. Measurements such as rolling or tipping behavior are treated as diagnostics that depend on the object's materials and shape. The instructions call for examining geometry and inspection images together.

Controller changes are compared with the existing version under matched test conditions. Shared random conditions reduce differences between trials; a failure can be replayed from the point it began before a full test sweep. The sim-to-real instructions reject changes that improve task results while worsening safety metrics. A single successful simulation run is insufficient evidence of improvement.

Two monitors show robot-arm simulations with different block placements, beside a physical bowl and blocks
AI-generated conceptual image: two scenes with different block placements represent testing across conditions · AI Ping

Preparing hardware still includes physical work by a person. The SO-101 instructions discover the connection port and cameras, then check calibration stored in the motors. Valid calibration is reused. When new calibration is needed, a person moves the arm through its joint ranges with motor torque off, producing settings for subsequent control.

Powered motion requires explicit human approval. Software preparation, read-only checks and hardware-free dry runs can proceed first. The SO-101 bring-up sequence includes checking the abort command, completing a test movement, returning to the starting pose and disabling motor torque.

Visual alignment before grasping repeatedly takes fresh observations. Its instructions pair each new camera frame with the actual joint position and wait for a commanded movement to finish before capturing the next image. If the target is lost, the arm widens the view and searches again. Failure to converge within the iteration limit triggers renewed localization and planning. The alignment helper uses CPU-based image processing, with camera calibration and the camera’s position and orientation relative to the arm as inputs.

A red emergency-stop button stands before a folded robot arm and camera, with an operator hand beside a keyboard
AI-generated conceptual image: an operator hand and emergency-stop button represent human supervision of physical hardware · AI Ping

Grasp verification is also hardware-specific. The FR3 Duo rules require evidence beyond how far the gripper has closed: its movement must be corroborated by a load change or views from multiple cameras. Before motion, the workflow checks that camera images are recent and have not frozen.

The plugin's code is available under Apache 2.0. Third-party software, robot models and external services such as the Gemini API retain their providers' separate terms.

The documentation assigns responsibility for safe physical operation to the user. FR3 Duo rules require idle actuators before anyone touches the equipment, cables or workspace, with a physical emergency stop within the operator's reach. Moving or bumping the arms, cameras or tabletop requires repeating preflight and camera checks.

Sources