OpenAI's robotics team has been disbanded once again, and GPT-6 Astra can directly drive robots
OpenAI has once again disbanded its robotics team. The former team has been merged into the core research team and now reports to Mark Chen, while the applied robotics and hardware teams are overseen by President Greg Brockman. This comes only five months after Altman's May post recruiting staff. The report states that in NVIDIA's simulated robotic arm test, the agent driven by GPT-6 Astra achieved a success rate of 74.8%, higher than the 51% achieved by the best specialized robotics model; in the same test, Claude Fable Astra's success rate was more than twice that of GPT-6 Astra. The article also mentions that running Astra on a dual-arm robot for one cycle at the highest configuration costs $12.5, and that OpenAI lost $10.9 billion in its most recent quarter.
Why read this
AI score 60
The original article places the team's disbandment alongside test results showing a general-purpose model directly controlling a robot, so readers can observe the changes taking place along the "robot brain" approach.
Topics
Body · AI translation
Incredibly, OpenAI has actually wiped out its own robotics team in one fell swoop!
According to reports, OpenAI has once again disbanded its robotics team, just five months after Altman personally posted a recruitment call in May.
The original team will be merged into the core research team, reporting to Mark Chen.
Meanwhile, the applied robotics and hardware team will be overseen by OpenAI President Greg Brockman.
Superintelligence Has Taken the Robotics Team's Jobs
At the Austin Conference on Robot Learning (CoRL) these past few days, you'll witness a bizarre sight—
At booth 208, there is only an empty table. Although the conference app and map both mark OpenAI, there are no robots on site.
Standing beside the booth are a few OpenAI people, awkwardly smiling at everyone they meet.
What's going on?
Those who know the truth were stunned—just this week, their robotics team was disbanded on the spot!
Is it because OpenAI couldn't keep going? Not at all.
But what they've done has left the entire robotics world feeling uneasy—now, there's simply no need to build brains for robots; GPT-6 Astra is the brain!
The industry knows that with the release of GPT-6 Astra, people are anxious: this is going to eat the entire robot brain industry.
If you want it to pour a glass of water, someone has to teach it hand by hand, action by action, using real data to specially train a "dedicated brain" for it.
But if you switch to another task, like tightening a screw? The robot immediately forgets everything.
But the current GPT-6 Astra has never trained on a single robot, yet it can directly pilot a mech.
Researchers gave it only three things: the camera feed, the robot's own state data, and a cable for the control interface.
Next, it's time to witness a miracle!
Astra is like a pro gamer who has watched countless walkthrough videos, starting at level 0 and immediately playing like a max-level account: it looks at the screen itself, breaks down the steps itself, and issues commands itself.
In a simulated robotic arm test run by Nvidia, the agent driven by Astra pushed the success rate straight to 74.8%.
In stark contrast, those "dedicated robot models" that have trained hard on real machines since childhood only reached a maximum of 51%.
In the face of a general top-tier large model, dedicated models seem utterly unnecessary.
Anthropic Crashes the Party: Claude's Success Rate Is More Than Twice That of GPT-6
Superintelligence isn't exclusive to OpenAI. In simulated tests, GPT-6 Astra's success rate at operating robots was less than half that of Claude Fable Astra.
And five days ago, Anthropic also showed off a "never eaten pork but seen a pig run" move.
In the official video for Fable 5.2, it controls two robotic arms, solving a Rubik's cube, unscrewing a bottle cap, and assembling gears.
It even drew a robot with a pen.
Moreover, the official announcement: Fable 5.2 has never been trained on these tasks.
Of course, the video is not in real time—solving the Rubik's cube alone took 40 minutes.
You should know that just four months ago, Claude was still a bit dim-witted.
At the time, it was controlling a robot dog walking, saw a trash can ahead, and confidently said, "The can is on the left, walk with confidence."
As a result—clang—the dog's leg got stuck in the can, and it dragged the can along for over two meters.
It's quite a sigh-inducing story: this is already the second time OpenAI has kicked its robotics team out the door.
The first time was in 2020.
At the time, Altman felt that building robots to collect data was too slow, and that LLMs were the true path to AGI.
Two years later, ChatGPT indeed burst onto the scene and shattered the world.
By 2024, they felt the large models were about ready, so they reassembled the robotics team and even partnered with humanoid robot company Figure.
As a result, by February 2025, the two sides fell out, so OpenAI decided to go it alone and declared: it would produce its first public demonstration in 2027.
But before the year is even out, the team is simply gone!
The first disbandment was because robots couldn't learn; the second is because the large model is too powerful, instantly becoming a max-level boss and taking over all the robots' work.
The brain belongs to superintelligence—what's left for robotics companies?
Once this news broke, the entire robotics industry was dumbfounded.
Over the past decade, capital has poured in tens of billions, and Figure's valuation has been hyped to $40 billion.
Everyone was betting on one thing: robots must have a specially customized brain.
As a CMU professor put it: in OpenAI's eyes, robotics "is no longer a special problem"; like tools for coding, law, and finance, it is just a downstream application of general intelligence.
In the past, if the bottle moved, the light dimmed, or the hand slipped—which of these wasn't a problem requiring special training?
Now, general models rely on powerful spatial perception and logical reasoning to improvise on the spot.
So what should those robotics companies do?
An MIT professor pointed out a clear path—
First, the models all run in the cloud; if the robot wants to move a finger, it still has to wait for the network.
If the network lags for a moment, the cup is already shattered.
Tsinghua and other institutions recently tested it: connecting Astra to a dual-arm robot and running a full loop at the highest configuration costs $12.5 (nearly 90 yuan)! Who can afford that?
And on real machines, understanding instructions is easy, but to control force and coordinate two hands subtly, you still need a specially trained small model.
Actually, Altman's move is understandable.
You should know that in the most recent quarter, OpenAI lost $10.9 billion.
Previously, Altman cut the not-so-profitable Sora project, and now he's disbanding the robotics team.
To put it plainly, OpenAI's wallet is tight too—even the landlord's family has no surplus grain.
Looking back, robotics scientists toiled for decades just to tailor-make the ultimate brain for robots.
Who could have imagined that what they finally got was a brain that never once worried about robots—truly a "bitter lesson"~
It has read all the text humanity has ever written and seen the entire world humanity has captured. Then one day, it reached out to the physical world.
Perhaps in the future, robots really won't need their own brains.
They are merely countless "figures" of superintelligence in this world, ready to be commanded at will.
When a whale falls, all things thrive. Who will the vacated track fall to?
References:
Editors: Aeneas, David
This article is from the WeChat public account "Xinzhiyuan" (ID: AI_era), author: Revelations of ASI, published by 36Kr with authorization.