Google releases Gemini 4 Argon for robot control, reports 87% on ERQA without helper tools

Google has released Gemini 4 Argon, a model it frames as a general-purpose physical-reasoning system for robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement, rather than as a single end-effector policy, according to a report by InfoQ.

Availability and pricing

Argon is available to Google AI Ultra subscribers and through the paid Gemini API. The model id in the Gemini API models docs is gemini-4-argon. Per-token pricing and context window size are not yet disclosed.

Benchmark claim

Google reports that Argon scores 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning, without helper tools. Google says this result leads other frontier models on that test. The figure is Google's own and was not independently verified at the time of the report.

What Google demonstrates

Google's published prompts cover several workflows:

• Legged robot control: Argon directs a Boston Dynamics Spot through a foot-pedal task after practice in a MuJoCo twin.
• Arm control: it steers a simpler arm using visual coordinates and HTTP motion primitives.
• World model loop: it closes a loop with the Gemini Omni video world model.
• Real2Sim: examples rebuild MuJoCo scenes from teleoperation video.
• Labeling: prompts annotate demonstration footage for behavior cloning.
• Policy improvement: Google reports synthesizing domain-randomized MuJoCo episodes from one RGB photo to fine-tune a vision-language-action (VLA) policy. It cites success on the corresponding real setup rising from 56% to 75%.

Caveats

The robotics post relies on long, task-specific prompts, coordinate grids, motion primitives and verifiers, so scaffolding is part of the system under test. Google does not publish per-demo failure rates beyond the single VLA case.

Early practitioner reaction centered on simulation fidelity. Davin YC Dong wrote on X that Google's examples tune simulation twins against contact events as well as visuals, adding: "a scene can look convincing while its objects behave incorrectly. For manipulation, matching what happens at contact matters as much as matching pixels."

Separately, a developer posting as pulpmatrix called for source discipline around Video-to-Sim-to-Real claims circulating the same week. The developer said the workflow is documented in Google's Gemini 4 Argon for Robotics dev blog (MuJoCo twin, 87% ERQA), but that a cornhole demonstration has no paper, repo or hit rate attached, and that real throws calibrated the robot first.

Related systems

InfoQ places Argon against specialized VLA systems such as OpenVLA-class policies, Physical Intelligence's π-family models, and NVIDIA GR00T with Isaac-oriented tooling. These optimize for action prediction under robot embodiments and typically rely on large teleoperation or simulation corpora. Scene-to-sim and generative world tools, including World Labs-style reconstructions, NVIDIA DreamDojo-style action video and Isaac Sim digital twins, emphasize environment synthesis or video prediction. They do not necessarily own the full control-and-QC loop shown in Google's prompts. Gemini Robotics-ER remains a related Google line for spatial perception and planning.

What this means

Argon's pitch is breadth: one model handling control, simulation construction, labeling and policy iteration. The evidence so far is a 87% ERQA score reported by Google and one quantified real-world result (56% to 75% on a VLA fine-tuning setup). Because the demos depend on heavy prompt scaffolding and lack per-demo failure rates, teams cannot yet separate the model's contribution from the harness. For architecture reviews, the open question is whether one frontier model plus prompt harnesses can displace a pipeline of separate VLA, simulator and labeling services. Contact-event accuracy in the simulation twins is the first thing practitioners say they would inspect.