Latest update
OpenAI reorganized its robotics team. Now GPT-6.2 Astra unfolds T-shirts and ties knots, doubling its robot success rate
First seen on X 2 days ago@gdb ♥ 9998 views
OpenAI's GPT-6.2 Astra became the first model to succeed at unfolding a T-shirt and tying a knot on a real-robot benchmark, more than doubling its success rate in two months.
Robotics evaluation company XDOF said on X on Nov 20 that GPT-6.2 Astra scored 41.2% on its Ember-0 benchmark, up from 19.6% for GPT-6 Astra in early September. Ember-0 has two robot arms attempt 16 tasks, such as unfolding a shirt and tying a knot, 30 times each, and averages the success rates. "Pick up bread with a tong" jumped from 10% to 90%, and XDOF said Astra is the first model to show any success on deformable-object tasks. "We expect Ember-0 to be saturated very soon," it wrote. The same day, Artificial Analysis's robot-arm benchmark AA-Robotics v0.1 put Astra first with 79 out of 100; it cleaned up a box in 92 arm commands, about 75% fewer than Claude Fable 5.2.
OpenAI President Greg Brockman quoted the result: "GPT-6.2 Astra is great at robotics." AI analyst Andrew Curran called it "the same unusual and unexpected jump in capabilities that Anthropic observed in Fable 5.2," and claimed the jump was why OpenAI restructured its dedicated robotics team. Fortune reported this month that OpenAI merged that team into its core research and applied organizations, betting on general-purpose models to drive robots. OpenAI staff said nothing was disbanded and the robotics work continues.
There are caveats. XDOF noted average prediction latency fell from 15.0 to 10.1 seconds but is still slow compared with other architectures. Robotics startup Waddle Labs said its agent Waddle-1 scored 81 on Robocurve versus 65 for GPT-6.2 Astra. Chat users have been harsher: The PrimeTime's "OpenAI is Cooked" passed 220,000 views in a day, and the comment "Last time I was this early OpenAI was an actual frontier AI company" got 418 likes.
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.
“GPT-6.2 Astra is great at robotics:”
