AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google wants Gemini 4 Argon to be the only AI model robots need

Collected Oct 15, 2026

Google has unveiled Gemini 4 Argon, a robotics-focused model it positions as a single system spanning the whole robotics pipeline: controlling robots, generating interactive 3D training environments, and evaluating training data. The model is available now in Google AI Studio and the development environment Antigravity, but only for Google AI Ultra subscribers, with access also through the paid tier of the Gemini API. The launch follows a global release of Gemini 4 Argon days earlier, and Google frames the work as another step toward "physical AGI," a goal it and its subsidiary DeepMind have pursued for years.

The clearest demonstration involves a Boston Dynamics Spot robot tossing soda cans into a step-on trash bin. Before touching the real pedal, Argon builds a virtual copy of the bin, and the robot practices balancing on three legs inside that simulation in MuJoCo. Only then does it attempt the task on real hardware, and two additional AI agents verify that the foot and gripper are positioned correctly before the can is dropped. The pattern — simulate first, then act — recurs across the examples.

Simulation generation is where the numbers get interesting. From a single photo of a table with party props, the model built a virtual replica including a soft plush cap, came up with 77 tasks ranging from righting toppled objects to clearing the table, and generated more than 4,500 training episodes for a control model. Tested on the real robot, success rate rose from 56 to 75 percent and execution time fell from 35 to 24 seconds. Argon also works from other inputs: a recording of a teleoperated robot yields a simulation designed to look identical and replicate fingertip force readings, a first-person video of a human grasping maps hand and finger motion onto a virtual five-fingered humanoid, and even a ballpoint pen sketch becomes a simulation — one drawing of a humanoid knocking down a box pyramid with a ball turns into a throwing motion that topples all six boxes without the robot falling over.

One demo plays out entirely inside an AI-generated video. Gemini Omni, a video world model that generates clips showing how a scene changes, takes timestamped instructions from Argon — which acts as director — describing what hands should do, such as sliding a Scrabble tile to the center of a table. Argon checks individual frames and adjusts the next command; after 33 steps, five scattered letter tiles are arranged to spell "ARGON." In a separate robot-arm demo, Argon placed a Lego brick into a wooden box and improved over attempts: the first run took just under a minute, the second 23 seconds though the brick clipped the box edge, and the third landed dead center in about 17 seconds. Across three iterations, execution time drops from 58.9 to 17.4 seconds. Google also wants Argon in research: given one prompt per paper, it recreates published robotics experiments as browser-run interactive simulations, with 68 such sandboxes built so far, including robot soccer and a humanoid on a skateboard.

The benchmark picture is mixed, which is worth noting. ERQA (Embodied Reasoning Question Answering) tests how well models understand and solve physical and spatial problems in 3D environments, using 400 questions without helper tools so models must rely on reasoning alone. Google reports Argon at 87 percent, ahead of all other frontier models; runner-up Claude Opus 5.5 reached 82.3 percent, GPT-6 Astra followed at 81 percent, and Claude Fable 5.1 trailed at 74 percent. A day later, Artificial Analysis published AA-Robotics, in which language models directly control a real robot arm across eight tasks. There Argon placed third with 55 out of 100 points, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65. The gap between a reasoning benchmark it leads and a direct-control benchmark it does not is a useful signal for anyone weighing claims of a single model covering the pipeline.

Argon is also meant to evaluate training data quality. In a video just over two minutes long showing a dual-arm robot folding an origami fox, it breaks the sequence into sub-steps and flags pauses and botched grips, and those segments are weighted less during training. In another video of a human sorting rings, Argon traced where the task progresses, stalls, or goes backward. A more playful demo involves the classic twelve-ball weighing puzzle: Argon built a simulation with two robot arms and a balance scale, reading the scale's tilt from camera images alone, and finds the odd ball in three weighings or fewer.

The published example prompts put some claims in perspective. For righting an overturned cup, described as zero-shot, Argon receives a camera image with a centimeter grid overlay and pre-built commands like "grasp" or "lower until contact." The simulations come with long prohibition lists to stop the model taking shortcuts such as copying measured sensor values. Google's blog post does not share failure rates for individual demos, and the 75 percent success rate lacks any mention of how many total attempts it is based on. This suggests the human scaffolding in the prompts is doing real work, and that "zero-shot" here means something narrower than a lay reader might assume.

For context, robot training in virtual environments is not new. The startup World Labs has shown how real scenes can become virtual training worlds, and Nvidia's DreamDojo simulates robot actions as videos. Google's differentiator, as it presents it, is ambition: handling simulation, control, and evaluation with the same model. In practice, the trade-off of one model across all three roles is that strengths in reasoning may not transfer cleanly to low-level control, exactly the split the two benchmarks hint at. A likely limitation is that the prompt engineering required per task may not scale as neatly as the demos imply.

Why it matters: developers and product teams working on robotics now have a paid, subscription-gated option that could replace several separate tools for simulation, control and data review, potentially shortening the loop from task idea to real-world test. The mixed benchmark results and the heavy prompt scaffolding mean teams should validate Argon on their own hardware and tasks rather than assuming the headline ERQA lead carries over to direct control, which is inference from the reported numbers. Research groups, meanwhile, gain a way to recreate papers as runnable browser simulations, though the durability of that capability is unproven.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Google's Gemini 4 Argon controls robots, generates simulations from photos or sketches, and evaluates training videos. On Google's own ERQA benchmark, it scores 87 percent. But the published example prompts show that the demos still need a lot of human guidance. The article Google wants Gemini 4 Argon to be the only AI model robots need appeared first on The Decoder .