Google's Gemini 4 Argon Targets Robotics With 87% Score on ERQA Benchmark

A benchmark claim and four areas of use

In a post on the Google Developers Blog, Principal Software Engineer Kanishka Rao describes Gemini 4 Argon as a model that unlocks new workflows across the robotics stack and speeds progress toward what Google calls physical AGI. As evidence, the company points to the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to understand and solve physical, spatial problems in 3D environments. Google states that Argon reaches 87% using pure reasoning, ahead of every other frontier model.

The post shares representative prompts in 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 detailed understanding of dexterity, dense value functions and progress estimation.
• Research and self-improvement: action improvement via simulation and iterative self-improvement on hardware.

Robot control examples

The first example has Argon controlling a Boston Dynamics Spot quadruped, described as having 12-DoF legs and a 6-DoF arm. The robot learns a new skill: stepping on a foot-pedal trash bin to throw away soda cans. Before stepping on the physical pedal, Argon builds an articulated MuJoCo digital twin of the bin and practices its moves there. The prompt asks Spot to verify through its front cameras that its front-right foot target is centered on the pedal, and to use a 2-agent verification swarm, a foot-verifier and a hand-verifier, to check both the pedal step and the 3D hand-in-bin alignment before releasing the can.

The second example is zero-shot control of a simple robotic arm that must place a fallen mug upright and put a crumpled napkin inside. According to the post, Argon reasons with a visual coordinate harness and action primitives, completing the task over several turns. The prompt supplies a calibrated wrist camera image with a projected 2 cm world coordinate grid, plus an observation file. It asks the model to answer each turn with a one-sentence observation followed by a curl command. The primitives include look_at, grasp, place, goto, move_by and move_until_contact, and the prompt lets the model choose how many visual checks to make, from 2 to 4 calls up to 8 to 20 for fine visual servoing.

The third example uses the Gemini Omni World Model. Argon steers robots there to rearrange five scattered Scrabble tiles (G, R, A, N, O) into a row spelling A-R-G-O-N. Rather than relying on one static prompt, Argon works in a closed visual loop: it writes timestamped action prompts, inspects intermediate rollout keyframes and refines its prompts. The demonstration is described as a 33-step steered video rollout with dexterous hands.

Real2Sim: building digital twins from robot data

In the Real2Sim section, Argon watches a real robot teleoperation episode and develops a digital twin. According to the post, it obtains or creates simulation assets that look and behave like the real objects when touched by dexterous hands. It then tunes the assets and scenes in an iterative visual-reasoning loop so that simulated visuals and contact events match the recorded real data.

The prompt sets two verification goals. The first is a visual match between MuJoCo renders from the head_left, left_wrist and right_wrist cameras and the real observations. The second is a physical match: simulated contact force and torque on each Sharpa fingertip must agree with the logged tactile_force in timing, magnitude and direction.

The rules limit what the model may change. Tunable elements are static object assets (geometry, collision shapes, mass, inertia, friction and joints, plus initial poses at t = 0) and offline-planned robot commands, altered only in frames where the real fingertips are near an object. Each command must stay within 0.55 rad for arm joints and 0.65 rad for hand joints of the logged teleop trajectory. The prompt also forbids "sim cheating": task objects may move only through real MuJoCo contact with the robot hands and the environment.

What it means for robotics developers

The post is a showcase of prompts rather than a full technical report, and the performance claims come from Google itself. The excerpt of the source provided does not include pricing, availability or release details for Gemini 4 Argon. What it does show is the intended role of the model: a reasoning layer that writes code against robot APIs, builds and checks simulations, and steers world models, with the same system covering control, simulation and data labeling. For robotics teams, the examples indicate that Google sees a general-purpose model, rather than separate specialized tools, as the way to connect these steps.