Gemini 4 Argon: Google Promises One Model for the Entire Robotics Stack — but the Numbers Are Company-Reported
Google launched Gemini 4 Argon on October 13, 2026 as a general physical reasoning model intended to cover the entire robotics stack, from control to simulation. But behind the marketing lie company-reported numbers without independent verification, and the coverage reveals that the demos still require detailed human guidance.
The News
On October 13, 2026, Google published a post about Gemini 4 Argon for robotics on its Developers Blog, written by principal software engineer Kanishka Rao (Google Developers Blog, source 04b9f4e2). Argon is presented there not as yet another robot model, but as a general physical reasoning model that can be applied "across the entire robotics lifecycle."
Google's own summary is ambitious: according to the company, Gemini 4 Argon marks "a step change in physical understanding, code generation, and visual-semantic reasoning" — and will "accelerate progress toward physical AGI" (Google Developers Blog, 04b9f4e2). It is worth noting that this is Google's own framing, not a documented capability.
Google positions Argon around four claimed use cases: robot control, Real2Sim (building simulations of the physical world), physical understanding, and research and self-improvement (Google Developers Blog, 04b9f4e2).
How It Works: Simulation First, Physics After
The core of the Argon demonstrations is a pattern in which the model does the cognitive work before a robot does the physical work: Argon builds its own simulation, trains or plans in it, and then transfers the result to a real robot.
The most concrete example from Google's own post is a Boston Dynamics Spot quadruped, equipped with a 12-degree-of-freedom leg set and a 6-degree-of-freedom arm, learning a new skill: stepping on the pedal mechanism of a trash can and throwing soda cans into it. According to Google, Argon first creates an articulated MuJoCo digital twin of the pedal trash can itself, rehearses the motions there, and then carries out the task in the physical world (Google Developers Blog, 04b9f4e2).
The Real2Sim portion is shown in another example: from a single photo of a table with party supplies, the model built a virtual copy, including a soft plush hat — an object with deformable physics, which is notoriously difficult to model. The model then came up with 77 different tasks, from righting overturned objects to clearing the table, and generated over 4,500 training episodes for a control model (The Decoder, 80353af9).
The Numbers — and What They Actually Show
As a caution for any reader of AI news: nearly all of the results here are company-reported.
The most prominent claim is a score of 87 percent on ERQA, the Embodied Reasoning benchmark, "with pure reasoning" — which Google claims is better than all other frontier models (Google Developers Blog, 04b9f4e2). This is a Google claim, not an independently verified result.
The Real2Sim example has more detailed, but still company-reported, numbers (The Decoder, 80353af9):
- The success rate on the real robot rose, according to Google, from 56 to 75 percent after sim training.
- Completion time fell from 35 to 24 seconds.
- The basis was 77 tasks that Argon itself came up with, and over 4,500 generated training episodes.
The improvement is significant if the numbers hold — 75 percent success is far from reliable enough for most real-world tasks, but the jump from 56 percent shows the point of sim training: the model learns cheaply in simulation before trying expensively in the real world.
In a third demo, Argon reportedly directed Google's video-based world model Gemini Omni with timestamped instructions, with the goal of arranging five scattered letter tiles to spell "ARGON." After 33 steps, they were in place (The Decoder, 80353af9).
The Caveats: Human Help and Paid Access
This is where the picture becomes more nuanced. The Decoder's Jonathan Kemper wrote on October 15, 2026 that the demos "still require detailed human prompting and guidance" (The Decoder, 80353af9). That stands in noticeable contrast to the framing of a single autonomous model covering the entire lifecycle: in practice, Argon in the demonstrations is a powerful tool under human direction, not an autonomous robot brain.
Argon is also available only through Google's paid developer programs (The Decoder, 80353af9) — researchers outside the company therefore have limited ability to test the claims themselves.
The Cornhole Demo: Social Media as a Source
A fourth demo has circulated on social media in the days around the launch, but it is less well documented than the others. According to an Instagram post by Sander Wubben, who describes a Google demo, an Apptronik Apollo humanoid learned to play cornhole by watching a video and then writing its own MuJoCo simulation code. Google says, according to Wubben, that this video-to-simulation pipeline can cut weeks of manual modeling down to hours — but it is an internal demo, not a published study, and Google has not stated how many attempts it took or how long it lasted (Instagram, 06ecd72b).
Another Instagram post, from CTO ROBOTICS Media, claims that the robot afterwards hit "7 throws in a row" (Instagram, d78531c5). The two posts agree on basic details — the Apptronik Apollo, MuJoCo practice — but they are not independent confirmation of each other; both are secondary renditions of Google media, and neither specifies attempts or duration. The "7 in a row" figure should be treated as unverified. It is also unclear in the sources how the Apollo demo relates to the official blog post of October 13.
What Is Missing
The clearest feature of the Argon launch is what is absent: no peer-reviewed study, no technical report with methodological details, and no assessments from independent robotics researchers in the available sources. All of the numbers — 87 percent on ERQA, 56 to 75 percent success, the cornhole results — come from Google itself or from social media renditions of Google material.
That does not mean the claims are wrong. The Real2Sim approach Argon demonstrates — a model sees a physical scene, builds its own simulation copy, and trains a robot in it — follows an established research direction, and Google's demos are described concretely enough to take seriously. But the distance between "the model can do this when Google is steering it" and "the only model robots need" is large, and the open questions are many: How robust is photo-to-simulation outside a controlled tabletop scene? How much prompting is really required per task? And will the results hold up when someone outside Google is allowed to measure?