Google Gemini 4 Argon Scores 87% on ERQA and Advances Robotics
Gemini 4 Argon scores 87% on ERQA, turns video and sketches into robot simulations, and boosts real-world task success from 56% to 75% for developers.
Summary
On October 13, 2026, Principal Software Engineer Kanishka Rao introduced Google's Gemini 4 Argon, a general-purpose robotics model scoring 87% on Embodied Reasoning, or ERQA, above every other frontier model using pure reasoning. Spanning control, Real2Sim, physical understanding and self-improvement, Argon created a MuJoCo twin before directing Boston Dynamics Spot's 12-DoF legs and 6-DoF arm to pedal a bin and discard cans, used zero-shot control to right a mug and insert a napkin, and steered Gemini Omni through a 33-step rollout arranging Scrabble tiles into A, R, G, O, N.
Real2Sim reconstructed contact-calibrated twins from robot footage, retargeted egocentric video to a 76-DoF bimanual humanoid with 22-DoF hands in 500 Hz MuJoCo, and converted an Apollo 2 sketch into a 50 Hz, 140-step, 2.80-second throw that kept pelvis drift below 3 cm and knocked down 6/6 boxes, with Blender 4.5 output. Argon annotated a 136-second ALOHA 2 origami run into six subtasks and 17 atomic primitives, reweighting failures 0.15x to 0.25x; solved the 12-ball Franka FR3 puzzle, distinguishing one 270 g or 30 g ball among eleven 150 g balls within three weighings; and split a 60-second UMI ring-sorting episode into 28 value segments.
From one photo, Argon modeled nine props, defined 77 tasks and generated 4,500+ episodes, raising real-world success from 56% to 75% and cutting completion from 35 to 24 seconds. A 6-DoF SO-101 improved pick-and-place from 58.9 to 22.9 to 17.4 seconds. Argon also created 68 MuJoCo WebAssembly sandboxes, including five paper reproductions covering OP3, DeepLoco, ANYmal C, Whole-Body MPC Go2 and HUSKY G1, verified over 300+ Chromium frames. The gemini-4-argon model is available through paid Gemini API access, while Google AI Ultra subscribers also receive Google AI Studio and Google Antigravity access.
Positives
- Argon's 87% ERQA score exceeds every other frontier model evaluated using pure reasoning.
- More than 4,500 simulated episodes increased real-world robot success from 56% to 75% and reduced completion time from 35 to 24 seconds.
- Real2Sim converted robot footage, human video and a paper sketch into physically calibrated MuJoCo environments.
- SO-101 pick-and-place time fell from 58.9 seconds to 17.4 seconds across three self-improvement iterations.
- Sixty-eight MuJoCo WebAssembly sandboxes demonstrate Argon's ability to reproduce robotics research as interactive browser labs.
Risks & concerns
- No rival scores or independent validation accompany the reported 87% ERQA result.
- SO-101's second iteration grazed the box rim, showing that faster execution initially reduced placement precision.
- Gemini API access requires a paid tier, while Google AI Studio and Antigravity access requires Google AI Ultra.
- The demonstrations are representative prompts, with no fleet-scale deployment or long-term reliability results detailed.
- Physical AGI remains a stated goal rather than a demonstrated capability.