Google published a breakdown of Gemini 4 Argon on the Developers Blog on October 13, 2026 — a model for physical reasoning and robotics. On the ERQA benchmark, a test for understanding physical and spatial tasks in 3D, Argon scored 87% through pure reasoning, without external tools. The main part of the publication is the real2sim pipeline: from a single photo or even a sketch, the model builds a calibrated digital copy of the scene in the MuJoCo simulator, generates thousands of episodes on it for fine-tuning a VLA policy, and a real robot then confirms the improvement on actual hardware. There is no finished product with an API in the material: for now, this is a breakdown with prompts and video demos.

image

What happened

In the breakdown, Google showed several demonstrations. Argon controlled Boston Dynamics' quadruped robot Spot with 12 degrees of freedom in its legs and 6 in its arm: the model first rehearsed the movement on an articulated digital copy of a pedal trash can in MuJoCo, and then the real Spot stepped on the pedal and threw out cans, and before that, two verifier agents, foot-verifier and hand-verifier, checked the action plan. A simple manipulator with a wrist camera, on Argon's commands, picked up a fallen mug and placed a crumpled napkin in it. From a hand-drawn sketch, the model built a simulation of a 76-degree-of-freedom humanoid: a 32-DoF body, two 22-DoF arms with five fingers, throwing a ball at a pyramid of six boxes. In a real scenario with two Franka FR3 manipulators, a single photo of a workspace turned into a calibrated MuJoCo scene with 77 tasks and over 4,500 domain-randomized episodes, on which a VLA model was fine-tuned: on real hardware, the success rate increased from 56% to 75%, and the task was accelerated from 35 to 24 seconds — without additional physical robot time. Separately, the automatic annotation of a 136-second ALOHA 2 recording, where the robot folds a paper fox, into 6 semantic subtasks and 17 atomic movements with failed attempt markers was shown; on the SO-101 manipulator, in three iterations, the task with a LEGO cube was accelerated from 58.9 to 17.4 seconds.

Context

Collecting real data is a historical bottleneck in robotics: VLA policies, which control robot actions based on observations and commands, are fine-tuned on episodes that a physical robot must perform itself, and each such episode costs robot time. The real2sim approach reverses the sequence: first, a digital copy of the scene is created, in which task variations are generated — in the breakdown, such episodes are called domain-randomized — then the policy is fine-tuned in simulation, and only the final check is carried out on real hardware. MuJoCo is a physics simulator long used in robotics, so the step that Google is demonstrating is not the simulation itself, but the automation of building a calibrated twin from a single photo or sketch and multi-agent verification before dangerous actions. ERQA, meanwhile, measures not control, but understanding: the model's ability to reason about physical and spatial tasks in 3D without recourse to external tools.

Why this matters for the industry

For teams building robotics and VLA systems, the post shows a transferable engineering pipeline: a photo of a workspace, a calibrated MuJoCo scene, thousands of domain-randomized episodes, a fine-tuned VLA policy with measurable verification on hardware. If similar improvements are reproduced on new scenes, the economics of policy preparation shifts from real robot time to computation in the simulator and scene calibration, and the shortage of real episodes ceases to be the main constraint. A second direct tool is the automatic annotation of accumulated recordings: dividing an operation into subtasks and atomic primitives with failure markers turns video archives into training datasets without manual annotation. Teams that already have physical robots and operation recordings can reproduce the combination of "video annotation, rehearsal in a MuJoCo twin, synthetic episodes for fine-tuning" on their existing stack, without waiting for Argon to become available. If Google opens the model via API, the robotics data prep category is likely to be formalized: automatic re-arrangement of robot videos into annotated datasets and generation of synthetic data for VLA policies.

Why this matters for users

You can read the full breakdown with prompts and watch the official video demos — Spot with a pedal trash can, ALOHA 2 with origami, and SO-101 with a LEGO cube — directly in the post on the Google Developers Blog. Argon cannot yet be used as a product: the source has no API, prices, or hardware requirements. Practitioners have access to the pattern itself: first, rehearsal of a skill in a MuJoCo twin of a real scene, then transfer to hardware with verification before the final action. For those who operate robots and accumulate operation recordings, the same scheme is useful from the other side — to annotate recordings automatically and fine-tune the policy in simulation, saving real robot time. For the rest, this is an early example of how the threshold for programming physical agents can be lowered to the level of "described or drew the task."

What is still unknown / limitations

The material is a blog post with prompts and videos, not a technical report: the source has no full description of the methodology, code, API, prices, or hardware requirements. The claim of superiority over any other leading model on ERQA belongs to Google and is not supported by a comparison table or methodology. For simulations built from a photo or sketch, the reconstruction accuracy metrics and the magnitude of the sim-to-real gap are not disclosed, so the word "calibrated" remains a declaration. The conclusion about the cheapening of training data is based on a single case — a workspace with two Franka FR3s, one task, a 19 percentage point increase — and has no baseline, for example a VLA model fine-tuned on an equal volume of real data. All demonstrations were filmed by Google itself, there are no independent checks in the source; even in the shown case, the success rate was only brought to 75%, meaning a quarter of attempts still end in failure.

Sources

Author

Look at AI, editorial team