Google DeepMind: Gemini 4 Argon scores 87% on ERQA embodied reasoning, targets full robotics stack

Google DeepMind says Gemini 4 Argon scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning, which the company claims outperforms every other frontier model. The result comes from a robotics-focused post published Oct. 13, 2026, by Kanishka Rao, a Principal Software Engineer at Google DeepMind.

ERQA tests a model's ability to understand and solve physical, spatial problems in 3D environments. DeepMind has not published a comparison table in the post, so the claim of leadership over other frontier models is unverified by third parties.

Four application areas

DeepMind describes Argon as a general-purpose physical reasoning model and shows representative prompts across four areas:

• Robot control: mobile manipulation with MuJoCo practice, zero-shot robot control, and closed-loop steering of video world models.
• Real2Sim: reconstructing physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn paper sketches.
• Physical understanding: agentic multimodal hindsight labeling, including dexterity understanding, dense value functions, and progress estimation.
• Research and self-improvement: action improvement via simulation and iterative self-improvement on hardware.

Demonstrations

Spot mobile manipulation. Argon controls a Boston Dynamics Spot quadruped (12-DoF legs plus a 6-DoF arm) to learn a foot-pedal trash bin task. The model builds an articulated MuJoCo digital twin of the bin, practices a 3-leg tripod balance in simulation, then performs the step on the physical pedal. The prompt instructs a 2-agent verification swarm (foot-verifier and hand-verifier) to check foot placement and 3D hand-in-bin alignment before release.

Zero-shot arm control. Argon uprights a fallen mug and places a crumpled napkin inside using a simple robotic arm. The model works over several turns, reading a calibrated wrist camera image annotated with a 2 cm world-coordinate grid. It emits code-as-policies primitives (look_at, grasp, place, place_until_contact, goto, move_by, move_until_contact) as HTTP calls to a local planner. The prompt lets the model choose its own call granularity, from 2–4 calls for coarse execution to 8–20 for fine visual servoing.

World model steering. Argon steers the Gemini Omni World Model in a closed visual loop to arrange five Scrabble tiles into the word A-R-G-O-N. It writes timestamped action prompts, inspects intermediate rollout keyframes, and refines its instructions. DeepMind shows a 33-step steered rollout with dexterous hands.

Real2Sim. Given a real robot teleoperation episode, Argon builds a MuJoCo scene of the robot and reconstructed task objects. It tunes assets and scenes in an iterative visual-reasoning loop until simulated visuals and contact events match the recorded data.

What was not disclosed

The post does not state Argon's context window, parameter count, pricing per 1M tokens, training cutoff, or API availability. It gives no benchmark scores beyond ERQA. The excerpt available for this report was truncated before the sections on physical understanding and self-improvement, so details of those demonstrations are not covered here.

What this means

The notable shift is architectural rather than numerical. DeepMind is presenting one general model as the reasoning layer across the robotics pipeline: writing control code, building simulation twins, labeling data, and prompting a separate world model. Each of those steps has traditionally needed a specialized system.

The Real2Sim and hindsight-labeling use cases may matter more in practice than the 87% ERQA figure. Calibrated simulation twins and dense progress estimates reduce the cost of generating robot training data, a major bottleneck. The demos are curated, single-task examples, and the post gives no success rates or failure analysis. Until DeepMind publishes full benchmark comparisons and access details, the lead over other frontier models remains a company claim.