Blog
Astra Picked Up a Robot Arm With Zero Training — Then Hit the Same Wall as Everyone
GPT-6 AstraRobot armsSep 2026
7 min read
On Sep 3, OpenAI released GPT-6 Astra — a model for computer use, coding, and office work. Within days, robotics testers did the obvious experiment: skip the office, plug it into robot arms, give it no training, and see what happens. What happened over the next two weeks is the clearest picture yet of where robots stand — and it’s worth telling in order.
19 Out of 20
The first test came from RoboCurve, an independent lab that grades AI models on real hardware instead of demo videos. Same two arms, same two jobs, three models. Job one: pick up a block, drop it in a bowl. Astra did it 19 times out of 20. Claude Fable 5.1 managed 8. The older Fable 5 managed 1.
And it did the job faster and cheaper — two and a half minutes per run instead of seven, under a dollar instead of two. The model looks through three cameras, decides where the gripper should go, and a standard controller moves the joints. No training, no tuning. For simple pick-and-place, the chat model beat the robotics specialists on their own ground.
95%Zero trainingPick and place
The Paintbrush
Then came the demo nobody asked for. A 20-year-old roboticist at OpenAI handed Astra an ordinary arm, a paintbrush, and a camera, and told it to paint the Golden Gate Bridge. No joint programming, no trajectories written by hand. The model figured out how to move the brush, watched its own work through the camera, and corrected its strokes over and over until the painting emerged.
That demo travelled further than any benchmark table, because anyone can read it: a general model learning a physical skill by watching itself. Amazon scientist Zeeshan Zia put the industry’s nervous thought into words — maybe the “OpenAI of robotics” is just OpenAI.
Zero-shotVisual feedback
The Groove Where It Stalls
But RoboCurve’s second job tells the other half. Pick up a round puzzle piece by its little knob and press it into a tight groove. Astra: 2 out of 20. Fable 5.1: exactly the same 2 out of 20. Every run ended the same way — the piece carried neatly over the groove, then a small miss on the final press.
The model can see fine. It can’t feel. Tiny errors stack up with each move, and without touch sensing and compliant joints, the last millimetre stays out of reach. Unitree’s CEO said it plainly: software lives in a perfect world, metal doesn’t. Separate MolmoSpaces tests confirm the pattern — Astra leads the generalists at 70.5%, while jobs needing real contact still belong to dedicated controllers and force feedback.
Precision wallTouch sensing
The Fight It Started
A 95% score with zero training is an attack on an entire business model. Specialist labs like Physical Intelligence and Skild raised hundreds of millions to build robot-only brains; if a general model does the job out of the box, what are they selling? Clone Robotics‘ CEO fired back that models go stale in six months while hardware moats last — actuators, sensors, and data from real deployments.
Both sides then reached for hardware. Figure AI had already walked away from OpenAI to build its own Helix brain. Altman confirmed OpenAI will “definitely” build its own humanoid. NVIDIA bought Hugging Face for $12.9 billion to own the simulation-to-model road. The scoreboard said software; the money says bodies.
HumanoidsOpenAIFigure
What Comes Next
Follow the story forward and the shape is clear. Simple jobs — sorting, kitting, moving boxes — get zero-training pilots within the year, because the model side is done and only integration remains. Precision jobs stay on specialised force-controlled tooling until touch sensing catches up. And the way you buy robots changes: audited success numbers on your parts replace demo videos, because RoboCurve just showed everyone how grading is done.
For Indian plants, that means two moves. Automate the simple jobs now with arms and AMRs that already exist — don’t wait for legs. And when any vendor pitches humanoids, ask for their audited trial numbers on your task, not their keynote. ERL commissions every cell that way: tested first, headlines never.
PilotsERL
Start With the Simple Jobs
Sorting, kitting, machine tending — ERL scopes pilots that prove out on your floor. Talk to us →