Google's Gemini 4 Argon scores 87% on embodied-reasoning benchmark

By Marcus Bennett

Google's Gemini 4 Argon scored 87% on the Embodied Reasoning (ERQA) benchmark — the highest mark on a test that measures a model's capacity to understand and solve physical, spatial problems in 3D environments.

What kind of model is Argon?

The post frames Argon as a general-purpose physical reasoning model applicable across the entire robotics lifecycle. It states the model "marks a step change in physical understanding, code generation, and visual-semantic reasoning," and adds that Argon "bridges coding, visual perception, physics simulation, and embodied robot control."

In practice, Argon reasons natively over multi-view video, 3D geometry, contact mechanics, and proprioceptive telemetry (joint-position and force feedback from a robot's internal sensors). The model interprets not only what objects occupy a workspace but how agents physically interact with them to accomplish whole-body dexterous tasks.

What did Google actually demonstrate?

The post lists nine representative workflows. Each pairs a natural-language prompt with attached inputs — images, video, JSON state, or paper PDFs — and asks Argon to execute the corresponding pipeline.

• Boston Dynamics Spot. Argon controlled a Spot quadruped — 12 degrees of freedom (DoF) in the legs plus a 6-DoF manipulator arm — to pick up two soda cans and dispose of them in a foot-pedal trash bin. Before touching hardware, the model built an articulated MuJoCo (a widely used open-source physics simulator) digital twin of the bin, then ran a two-agent verification swarm: a foot-verifier checking foot-target alignment and a hand-verifier checking 3D hand-in-bin alignment before release.
• Mug uprighting. Using a visual coordinate harness (an annotated wrist-camera overlay mapping pixels to centimeters) and action primitives, Argon righted a fallen mug and inserted a crumpled napkin across several turns of wrist-camera feedback.
• Scrabble rearrangement in Gemini Omni. The model steered a virtual robot inside Gemini Omni (Google's video-based world model) to gather scattered tiles G, R, A, N, O and arrange them to spell A-R-G-O-N. Argon wrote timestamped action prompts, inspected intermediate rollout keyframes, and refined instructions in a closed visual loop.
• Digital-twin reconstruction. From one teleoperation episode (a recorded session of a human remotely controlling the robot), Argon built a MuJoCo digital twin by generating or obtaining simulation assets matching the real object's geometry, mass, friction, and joints. The model then tuned those assets against logged tactile force data from Sharpa fingertips (high-resolution tactile sensors mounted on each finger pad).
• Human-to-humanoid retargeting. From an egocentric human video, Argon identified task-relevant objects, replaced bounding boxes with compound convex-decomposed meshes (collision geometry split into convex pieces for accurate physics in concave containers), and retargeted the hand trajectory onto a 76-DoF bimanual humanoid with two 22-DoF five-fingered hands. MuJoCo contact physics validated the retargeted motion at 500 Hz.
• Sketch-to-simulation. From a hand-drawn sketch of an Apollo 2 humanoid throwing a ball at a 3-2-1 pyramid of six boxes, Argon constructed a mujoco_warp scene (MuJoCo's GPU-accelerated simulator backend), synthesized a 50 Hz whole-body throwing trajectory of 140 steps over 2.80 seconds, verified 6/6 knockdown in closed-loop physics, and exported a Blender 4.5 Cycles (a path-tracing rendering engine) video.
• ALOHA 2 origami annotation. Given a 136-second overhead video of an ALOHA 2 robot folding a square sheet into an origami fox, Argon produced a four-level hindsight annotation (time-stamped labels generated after the fact): an episode goal (G1), six contiguous semantic subtasks (G2), 17 contact-level atomic primitives (G3), and automated quality-control reweighting that holds task progress flat during fumbles and downweights them to 0.15x–0.25x for behavior cloning (training a robot policy by imitating logged expert actions).
• 12-ball balance puzzle. The model constructed a bimanual Franka FR3 (a 7-DoF research robot arm) MuJoCo environment containing 12 visually identical pool balls — 11 at 150 grams, one at 270 grams or 30 grams — and an articulated Roberval two-pan balance scale (a balance mechanism using parallelogram linkages). A vision-only solver triangulated the balls from stereo RGB cameras and read scale tilt from stereo vertical pixel disparity, isolating the odd ball into the correct Heavy or Light tray within three weighings.
• Ring-sorting progress model. Across a 60-second bimanual ring-sorting episode from a human UMI (a handheld gripper for collecting manipulation demonstrations) dataset, Argon produced dense value-function predictions (numerical estimates of task-completion progress) and decomposed the trajectory into 28 segments with positive, neutral, and negative progress slopes.

What does Argon do for VLA training?

Argon can synthesize simulation data to fine-tune generalist vision-language-action (VLA) models — AI systems that map camera inputs directly to robot actions. From one RGB photo of a real tabletop, the model autonomously built a calibrated MuJoCo twin, defined a 77-task curriculum of pick-and-place, uprighting, directional pushing, mid-air handovers, scoop-and-push, and pipelined table-clearing tasks, and generated more than 4,500 domain-randomized episodes (training scenarios with randomized object positions, lighting, and physics). Real-world Success Rate on the resulting policy rose from 56% to 75%. Mean completion time fell from 35 seconds to 24 seconds.

How does Argon improve itself on hardware?

On a 6-DoF arm, Google demonstrated a three-stage self-improvement loop.

Stage 1 measures objects. No hardware moves. The model dry-runs a pick-and-place plan.

Stage 2 hovers the gripper above the brick. The wrist camera verifies alignment.

Stage 3 executes the task autonomously, diagnoses failed attempts from wrist and overhead footage, retries, then optimizes the successful motion to at least twice its initial execution speed within safety limits.

Can Argon reproduce published research?

The release describes a workflow that takes a single robotics paper and any released checkpoints (saved model weights from training) and ports the physics model and closed-loop controller into a self-contained MuJoCo WebAssembly browser lab (a physics simulator compiled to run in web browsers at near-native speed). Examples cited span 7-network hierarchical locomotion from DeepLoco (a hierarchical locomotion framework), 50 Hz PPO actors (Proximal Policy Optimization, a reinforcement-learning algorithm) from HUSKY, and Web Worker iLQR (iterative Linear Quadratic Regulator, an optimization-based controller running on background browser threads) from Whole-Body MPC (model-predictive control, computing optimal actions over a rolling time horizon). Headless Chromium (a browser-automation framework running Chrome without a visible window) verifies each lab with zero console errors and non-zero progress over 300+ frames.

Why does Google frame this as a milestone?

The post concludes that Argon "marks an important milestone accelerating the path toward physical AGI" (artificial general intelligence operating in physical environments). Google's stated goal is to "enable robotics that can work alongside humans to solve complex challenges."

When can developers access it?

Argon is available today through the Interactions API (Google's developer endpoint for invoking Gemini models) with day-one support across the google-genai SDKs (Google's generative-AI software development kits). The release points developers to gemini-4-argon in the API for goal runs and multi-agent orchestration (coordinated operation of several AI agents on shared tasks).

What is the broader significance?

The 87% ERQA score, the breadth of hardware targets — Boston Dynamics Spot, ALOHA 2, Franka FR3, 6-DoF arms, 22-DoF humanoids — and the 19-percentage-point lift in real-world Success Rate from synthesized data together suggest that foundation-model reasoning has become a measurable contributor to robotic policy training, not merely a perception layer. Whether those gains hold outside the lab — across non-toy environments, off-distribution objects, and long-horizon mobile manipulation — will determine whether today's release anchors a sustained robotics stack or remains a single benchmark victory.