Argon Robotics: The Numbers Behind the 7 Straight Makes
Teaching robots is slow and expensive — people hand-code every motion and test hundreds of times on real hardware. Argon flipped the order: watch a human on video, practice thousands of times in simulation, then move to the real robot. Success climbed from 56% to 75%. But some of the bragging numbers are exam scores you should never mix with field ability. Below, I split what's real from what's dressed up.
The short version
Flow — video → simulation → real robot
Results — 4,500+ virtual trials, success 56% → 75%
Careful — ERQA 87% is a multiple-choice exam score; total cornhole attempts undisclosed
Video → simulation → real robot
- It watches a human play cornhole (bean-bag toss) and works out the coordinates and joint movements.
- It writes the simulation code itself.
- The humanoid Apollo, loaded with those motions, made 7 in a row.
Sources: AI Matters · AI Times
▶ Watch
ERQA 87%: what's behind the number
- It scored 87% on ERQA, an embodied-reasoning benchmark — higher, Google says, than the other frontier models it compared against.
- Here's the catch. ERQA is 400 multiple-choice questions using images and text to test 3D space and physics reasoning. It never checks whether a robot actually completes a real task. Exam scores and field ability are separate things.
Sources: AI Times · AI Matters
Where robot development changes
- Making robot motions: it watches human video, computes the motions the robot should copy, and writes the simulation code too. Humans no longer hand-code every coordinate — the video does it.
- Building virtual test scenes: it sets up experiments in the MuJoCo physics simulator, reviews failed attempts, and revises the robot's motions. Failures stop costing you broken hardware.
- Video conversion: first-person footage a person shot gets converted into a virtual model of a two-armed humanoid. Footage you already have becomes training material as-is.
- Sketch conversion: a hand-drawn sketch becomes a virtual test scene. Scenes that are hard to describe in words can start from a single drawing.
- Where: Google AI Studio and Antigravity for Ultra subscribers; developers via the paid Gemini API.
Source: AI Matters
The public demos
- Google + Apptronik: they taught the humanoid Apollo cornhole. After watching video and practicing in virtual space, the real robot adjusted its throws based on where the bags landed. A video of 7 straight makes was published — but the total number of attempts was never disclosed.
- Google's virtual experiments: 4,500+ runs in MuJoCo lifted real-task success from 56% to 75%. Simulation modeling that used to take weeks now takes hours, Google says.
- Boston Dynamics Spot: the quadruped Spot was also shown pressing a pedal to open a trash bin and tossing in two empty cans.
Sources: AI Matters · AI Times
What still needs checking
- ERQA 87% is a multiple-choice exam score. Keep it separate from real task-completion ability.
- With total cornhole attempts undisclosed, you can't compute a success rate from the 7 straight makes.
- Every case so far is a demo by Google and its partners. There are no third-party field validations yet.
- The numbers are good — field verification starts now.
Puzz's one-liner
The flow — watch, practice, move to the real robot — is the real story. But the comparison scores are exam-based, and success rates should wait until every attempt is public. Ribbit.