Google DeepMind has announced Gemini 4 Argon, a model the company describes as a step change in physical understanding, code generation, and visual-semantic reasoning. In a post on the Google DeepMind blog published October 13, 2026, the team behind the model said it unlocks new workflows across the robotics stack and accelerates progress toward physical AGI, with the stated goal of enabling robots that can work alongside humans to solve complex challenges.

The capability leap is most visible on the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to understand and solve physical, spatial problems in 3D environments. Using pure reasoning, Argon scores 87% on the benchmark, outperforming every other frontier model, according to DeepMind. Benchmarks like ERQA carry weight because frontier models have historically been measured chiefly on language and coding tasks, while reasoning about objects, space, and contact in 3D has remained a specialized challenge for robotics.

Positioned as a general-purpose physical reasoning model, Argon is designed to be applied across the entire robotics lifecycle. The announcement walks through representative prompts in four key areas: robot control, Real2Sim, physical understanding, and research and self-improvement. Each demonstration is accompanied by video of the model at work.

Robot control: from a quadruped to video world models

The robot control demonstrations cover mobile manipulation with MuJoCo practice, zero-shot robot control, and closed-loop steering of video world models.

In the first demonstration, Argon controls a Boston Dynamics Spot quadruped — a platform with 12-DoF legs and a 6-DoF arm — to learn a new skill: stepping on a foot-pedal trash bin to throw away soda cans. Before touching the physical pedal, Argon creates an articulated digital twin of the bin in MuJoCo, an open-source physics simulator widely used in robotics research, and practices its moves in simulation. It then steps on the real pedal and drops soda cans into the open bin.

In a zero-shot control demo, Argon operates a simple robotic arm to set a fallen mug upright and place a crumpled napkin inside it, reasoning with a visual coordinate harness and action primitives over several turns.

The third example extends control into video world models. Argon steers robots inside the Gemini Omni World Model to rearrange five scattered Scrabble tiles (G, R, A, N, O) into a clean row spelling A-R-G-O-N. Rather than issuing one static prompt, Argon steers in a closed visual loop: it writes timestamped action prompts, inspects intermediate rollout keyframes, and refines its prompts as the rollout progresses.

Real2Sim: from camera feeds and sketches to calibrated simulation

The Real2Sim demonstrations cover reconstructing physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn paper sketches. The technique addresses a practical reality of robotics: iterating in simulation is faster and safer than on physical hardware, but it only pays off when the simulated world faithfully matches the real one.

In the first, Argon views a real robot teleoperation episode and develops a digital twin, obtaining or creating simulation assets that look and behave like the real objects when the dexterous hands make contact. It tunes the assets and scenes in an iterative visual-reasoning loop until simulated visuals and contact events match those in the real recorded data.

Given an egocentric human video of a manipulation task, Argon identifies the task-relevant objects, replaces crude bounding boxes with compound convex-decomposed meshes for concave containers, and retargets the human hand trajectory onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands in MuJoCo contact physics.

The sketch-to-simulation example shows a developer snapping a photo of a rough pen-and-paper sketch — depicting an Apollo 2 humanoid throwing a ball along a parabolic arc to topple a 3-2-1 stack of boxes on a table — after which Argon automatically constructs the GPU-accelerated mujoco_warp scene, synthesizes a balanced 76-DoF whole-body and five-finger throw trajectory, verifies a 6/6 target knockdown in closed-loop physics, and exports a Blender 4.5 Cycles render.

Physical understanding: hindsight labeling, puzzle solving, and value functions

The physical understanding demonstrations cover agentic multimodal hindsight labeling, including detailed understanding of dexterity, dense value functions, and progress estimation.

DeepMind says Argon has precise physical understanding of dexterous and can provide detailed segmentation and text labeling from video, identifying poor manipulation segments such as pauses, unstable grasps, and fumbles. Given an overhead video of a bimanual ALOHA 2 robot folding a square sheet of paper into an origami fox over 136 seconds, Argon produces a time-aligned, four-level hindsight annotation hierarchy: an Episode Goal (G1), 6 contiguous Semantic Subtasks (G2), 17 contact-level Atomic Primitives (G3), and automated Quality Control (QC) reweighting that holds task progress flat during fumbles and retries.

To recreate the classic 12-ball balance-scale puzzle with robots, Argon builds a bimanual Franka FR3 MuJoCo environment containing 12 visually identical pool balls — one secretly heavier or lighter — and an articulated Roberval two-pan balance scale. It then writes a vision-only closed-loop solver that triangulates the balls from stereo RGB cameras, visually reads the scale tilt after each weighing, and isolates the odd ball into the Heavy (Red) or Light (Blue) tray within 3 weighings.

DeepMind also describes Argon as a general value function for robotics tasks. Evaluated on a 60-second bimanual ring-sorting episode from a human UMI dataset, Argon iteratively inspects synchronized head, left-gripper, and right-gripper camera feeds around contact transitions to generate dense value function and task progress predictions, decomposing the trajectory into 28 segments with positive, neutral, and negative progress slopes.

Research and self-improvement: from simulation data to hardware iteration

The research and self-improvement demonstrations cover action improvement via simulation and iterative self-improvement on hardware.

Argon can generate new data in simulation to improve the performance of generalist vision-language-action (VLA) models. For VLA models, which depend on large volumes of robot interaction data, simulation offers a way to scale data collection beyond what physical hardware allows. Starting from a single RGB photo of a real tabletop, it autonomously builds a calibrated MuJoCo simulation twin, defines new tasks, and synthesizes a closed-loop motion generator to collect more than 4,500 domain-randomized episodes in simulation. According to DeepMind, this boosted the real-world physical robot success rate from 56% to 75% and reduced completion time from 35 seconds to 24 seconds.

On a simple 6-DoF arm, Argon also iteratively self-improves both execution speed and placement accuracy on a pick-and-place task through trial and self-reflection.

The final demonstration covers one-prompt reproductions of robotics control papers in MuJoCo WebAssembly. With a single prompt per paper, Argon reads a robotics control manuscript and any released checkpoints, ports the physics model and closed-loop controller — from the 7-network hierarchical locomotion of DeepLoco and the 50 Hz PPO actors of HUSKY to the Web Worker iLQR of Whole-Body MPC — into self-contained MuJoCo WebAssembly browser labs, and verifies them in headless Chromium.

Advancing embodied intelligence

According to DeepMind, Argon bridges coding, visual perception, physics simulation, and embodied robot control. By reasoning natively over multi-view video, 3D geometry, contact mechanics, and proprioceptive telemetry, the model understands not just what objects are in a scene, but how they physically interact with agents to accomplish whole-body dexterous tasks. Taken together, the demonstrations span work that robotics teams typically distribute across separate tools and workflows — controller design, simulation calibration, dataset labeling, and paper reproduction.

"Gemini 4 Argon marks an important milestone accelerating the path toward physical AGI," the blog post states. "By building this core intelligence, our goal is to enable robotics that can work alongside humans to solve complex challenges."

Getting started

Gemini 4 Argon is available today across Google's developer platforms, giving developers a direct way to test the capabilities described above:

• Google AI Studio: available to Google AI Ultra subscribers via Google AI Studio.
• Gemini API: available with the paid tier of the Gemini API in AI Studio. Developers can invoke the model as gemini-4-argon in the Interactions API, with day-one support across the google-genai SDKs.
• Google Antigravity: available to Google AI Ultra subscribers. Developers can select Gemini 4 Argon (gemini-4-argon) as their active model in Google Antigravity for complex planning, /goal runs, and multi-agent orchestration.