Google's Gemini 4 Argon Robotics Post Reports 87% on ERQA, and Most Demos Run in Simulation

Google published a robotics write-up for Gemini 4 Argon on Oct. 13 that scores the model 87% on ERQA, a test of spatial and physical problem-solving, and lists more than a dozen demonstrations. Most of them are in simulation or are single runs on one robot. The Google Developers Blog post was written by Kanishka Rao, a principal software engineer. It describes gemini-4-argon as a general-purpose physical reasoning model.

The 87% ERQA figure is Google's own, and the post says it outperforms every other frontier model. It gives no scores for any competitor. ERQA is the only benchmark number in the post. The model's launch post, written by Koray Kavukcuoglu, repeats the same figure under Google's main Argon announcement.

What ran on hardware

The clearest hardware claim comes from the main announcement. Argon learned a cornhole throw from a video, built a simulation of it, and the skill transferred to Apptronik's Apollo humanoid, which sank seven consecutive shots, according to Google. The pipeline, which Google calls Video-to-Sim-to-Real, compresses weeks of simulation work into hours, it said.

The robotics post adds a Boston Dynamics Spot with a six-degree-of-freedom arm that learns to step on a foot-pedal bin and drop soda cans. Argon first builds an articulated MuJoCo twin to practice in. The post calls the prompt autonomous but does not say whether the full run was.

Boston Dynamics lists Spot's maximum payload at 14 kg and average runtime at about 90 minutes, and gives no arm specification or price on its product page. The Google post does not report how long the soda-can task took or how many attempts failed.

Demonstrations in the post, by type

Demo	Where it ran	Reported result
Sketch of Apollo 2 throwing a ball	Simulation	6 of 6 targets knocked down, under 3 cm pelvis drift
Simulation twin from one photo	Real robot, after fine-tuning	Success 56% to 75%, time 35s to 24s
SO-101 arm self-improvement	Real arm, stage 3 autonomous	Completion 58.9s, 22.9s, 17.4s
Odd-ball puzzle, dual Franka FR3	Simulation	Odd ball found in 3 weighings
Egocentric video to humanoid	Simulation at 500 Hz	76-DoF bimanual humanoid

The strongest number for anyone building robots is the second row. From a single RGB photo, Argon builds a simulation twin and collects more than 4,500 domain-randomized episodes for fine-tuning a vision-language-action model. The post reports real-world success rising from 56% to 75% and completion time falling from 35 seconds to 24 seconds. It does not name the task or the robot in that passage.

What the post leaves out

The origami-fox annotation demo uses an ALOHA 2 video of 136 seconds, and the post does not say whether the robot was teleoperated. The robot-view-to-simulation demo starts from a teleoperation episode, which the post does say. The post lists no limitations. Its own framing is that Argon is "an important milestone" toward physical AGI.

Availability is narrower than for the language model. Argon for robotics is in Google AI Studio and Antigravity for Google AI Ultra subscribers, and on the Gemini API paid tier through the Interactions API. The post lists no per-token price for the robotics use. The language-model pricing is $2 and $10 per million tokens until Nov. 30, 2026, per Google's announcement.

Where the money is going

Robot foundation models are drawing venture funding without matching deployment data. FieldAI has a $10 billion term sheet and General Intuition raised $220 million, both on the premise that a general model can steer many bodies.

Google's post is a developer preview, not a customer deployment. No customer is named, and Google has not said which of the demos, if any, are in production.

The next fixed date is Nov. 30, when Argon's introductory token price ends, according to Google.