Google's Gemini 4 Argon targets robotics with 87% on ERQA physical reasoning benchmark

What Google announced

Google has introduced Gemini 4 Argon, a robotics-oriented version of its frontier model. According to a post on the company's developers blog that surfaced on the front page of Hacker News, Argon delivers a "step change" in physical understanding, code generation and visual-semantic reasoning, and Google frames the work as progress toward "physical AGI." The post positions Argon as a general-purpose physical reasoning model that applies across the entire robotics lifecycle rather than to a single task.

The headline number is 87% on the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to understand and solve physical, spatial problems in 3D environments. Google says Argon reaches that score using pure reasoning and that it outperforms every other frontier model. As with any vendor benchmark, the result is Google's own claim, presented without third-party evaluation in the post.

Robot control, three ways

Google illustrates robot control with three demos of increasing abstraction.

In the first, Argon pilots a Boston Dynamics Spot quadruped — 12 degrees of freedom in the legs plus a 6-DoF arm — through a novel task: stepping on a foot-pedal trash bin so soda cans can be dropped inside. Before touching the physical pedal, the model constructs an articulated MuJoCo digital twin of the bin and rehearses the move in simulation, including the three-legged tripod balance required while the arm works. The published prompt also shows a "two-agent verification swarm": separate verifier agents must confirm that the foot is centered on the pedal and that the hand is aligned over the open bin before the robot releases the can.

The second demo is zero-shot control of a simple robot arm asked to stand a fallen mug upright and place a crumpled napkin inside it. Argon works from an annotated visual coordinate harness — a calibrated wrist camera carrying a 2 cm world-coordinate grid, a crosshair for the gripper's pinch center, and markers for jaw opening and object orientation — and emits action primitives such as look_at, grasp, place_until_contact and move_by across several closed-loop turns, re-inspecting the camera image after each step and recovering when a grasp fails.

The third demo moves into generative simulation: Argon steers robotic hands inside the Gemini Omni World Model to arrange five scattered Scrabble tiles into a row spelling A-R-G-O-N. Rather than issuing one static instruction, it writes timestamped action prompts, reviews intermediate keyframes from the rollout, and refines its prompts accordingly — a 33-step steered video rollout in the example Google shows.

Real2Sim and the rest of the stack

The second application area is Real2Sim: building simulation twins calibrated to the physical world from inputs as varied as robot camera feeds, egocentric human video or hand-drawn paper sketches. In the detailed example, Argon watches a recorded teleoperation episode and produces a digital twin, sourcing or creating simulation assets that match how real objects look and behave on contact, then iteratively tuning the scene until simulated visuals and contact events agree with the real recording.

The post also names two further areas with less accompanying detail. "Physical understanding" covers agentic multimodal hindsight labeling, fine-grained analysis of dexterity, dense value functions and progress estimation. "Research and self-improvement" covers improving actions in simulation and iterative self-improvement directly on hardware.

Why it matters

Robotics has trailed the rest of AI largely because physical reasoning, sim-to-real transfer and per-robot engineering resist scaling. Argon is an attempt to collapse that stack into one model: a system that builds its own simulation twins, practices in them, writes its own control code, verifies its actions with sub-agents and improves from its own hardware experience. The zero-shot mug demo is the clearest signal of what that could mean — robots handling tasks nobody explicitly trained them for. The caveats are just as clear: every number and demonstration comes from Google's own promotional post, and "physical AGI" is a framing rather than a measurable milestone. Whether Argon is a genuine step change or a well-rehearsed showcase will depend on independent evaluation and on whether the behaviors hold outside curated environments.