Google's Gemini 4 Argon Scores 87% on ERQA, Controls Spot

Google released Gemini 4 Argon on October 13, 2026, positioning it as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal video labeling, and iterative policy improvement. For teams assembling separate vision-language-action models, simulators, and labeling pipelines, the immediate question is whether one frontier model with structured prompt harnesses can replace three or four services.

Capabilities across the robotics lifecycle

In Google's developer blog, principal software engineer Kanishka Rao details four functional areas. In robot control, Argon directs a Boston Dynamics Spot quadruped—configured as 12-DoF legs plus a 6-DoF arm—through a foot-pedal trash-bin task. It first constructs an articulated MuJoCo digital twin of the bin for practice, then runs on hardware with a two-agent verification swarm: a foot-verifier and hand-verifier confirm the foot-pedal step and 3D hand-in-bin alignment before releasing the cans.

A second demo uses a visual coordinate harness with a 2 cm grid projected onto a wrist camera feed, issuing curl commands against a local HTTP endpoint and selecting primitives such as robot.grasp, robot.place_until_contact, and robot.move_until_contact. A third closes a visual loop around the Gemini Omni video world model, producing a 33-step rollout that rearranges five Scrabble tiles into A-R-G-O-N.

In Real2Sim, Argon reconstructs MuJoCo scenes from robot teleoperation video. Offline trajectory changes are limited to within 0.55 rad of logged arm-joint positions and 0.65 rad of logged hand-joint positions. After t=0, external forces, welds, and equality constraints are not allowed; objects move only through MuJoCo contact. A separate demo retargets egocentric human hand video onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands, running at 500 Hz (dt = 0.002 s) without kinematic weld constraints.

For policy improvement, Google reports synthesizing domain-randomized MuJoCo episodes from a single RGB photo to fine-tune a vision-language-action policy, citing a success-rate rise from 56% to 75% on the corresponding real setup.

Benchmark performance and availability

Google reports Argon scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning without auxiliary tools, and says that result leads all other frontier models on the test. The score and the claim that it leads are Google-reported and have not been independently verified. Argon is available to Google AI Ultra subscribers and through the paid Gemini API as gemini-4-argon. Google publishes no per-demonstration failure rates beyond the VLA fine-tuning case.

Competitive positioning

As InfoQ reported, early practitioner reaction focused on simulation fidelity. Developer Davin YC Dong commented on X that a scene can look convincing while its objects behave incorrectly, and that for manipulation matching contact events matters as much as matching pixels. That matters because Argon's Real2Sim prompt requires simulated fingertip contact force and torque to match logged tactile readings in timing, magnitude, and direction—enforced through iterative reasoning, not a physics solver tuned by domain experts.

AI Mastery analysis

The deepest question is whether prompt-harness scaffolding can be a production architecture. The demos depend on long task-specific prompts, coordinate grids, multi-agent verifiers, and timestamped action sequences. That is closer to pipeline architecture than a monolithic learned policy, and it carries the same risks: prompt brittleness under distribution shift, latency from closed-loop VLM calls, and missing failure rates outside the single VLA case.

The Real2Sim constraint set—joint deviation limits, no post-initialization model changes, contact forces sourced only from MuJoCo—is the most replicable specification in the release. Teams evaluating Argon against dedicated VLA pipelines should treat the 500 Hz retargeting demo as a benchmark for their own scene complexity, since the egocentric-video workflow assumes triangulatable camera intrinsics and convex-decomposable object geometry that may not hold in cluttered production environments. The architectural specificity behind these gains lies in the constraint system around the simulator rather than the model weights alone.

Argon suggests Google is treating the robotics stack as a prompt-engineering and scaffolding problem as much as a model-weight problem. If the ERQA score and the VLA fine-tuning result hold under third-party replication, teams maintaining separate labeling, simulation, and improvement services will face real pressure—but the complexity shifts to prompt engineering, a tradeoff production teams will need to price.

Sources

• Primary: [Gemini 4 Argon for Robotics

• Google Developers Blog](https://developers.googleblog.com/gemini-4-argon-for-robotics/)

• Reported by: InfoQ AI: Google Releases Gemini 4 Argon for Robotics Control, Simulation, and Policy Improvement