SI Agent News
DevelopingClaudeOpenAIGemini

OpenAI folds its robotics team again while Chinese firms back the robot control layer

OpenAI shut its robotics group for the second time in six years. Chinese robot makers are betting on the layer between big AI models and motors.

66.3%Astra's success rate directing robot hands in simulation

The short version

Why it matters

A top AI lab stepping back from dedicated robotics does not mean robots no longer need specialist software. The money may sit in the layer that turns an AI model's intent into fast, cheap, reliable movement, and that layer is where the Chinese bets are going.

Two shutdowns, two reasons

As tmtpost.com reported, Fortune revealed that OpenAI has dissolved its independent robotics team. Researchers move under the chief research officer, while applied robotics and hardware staff go under the president's applications division.

The first closure, six years ago, was blamed on a lack of training data. This time the stated reason is that frontier general models such as GPT-6 Astra can already handle robot tasks, so a separate group is unnecessary. The timetable for a first public humanoid demo has not changed.

What the evidence shows

OpenAI's confidence leans on an Nvidia robotics lab report. With no robot-specific training, Astra reached 66.3% on a simulated set of tabletop pick-and-place tasks. Inside a persistent coding session that keeps history and notes, it reached 74.8%.

But the result depended on what the model was asked to output. When it named where the robot's hand should go, it scored 66.3%. When it had to set joint positions, it dropped to 50.2%, and with low-level trajectories to 38.8%. In plain terms, the model is good at deciding where the hand goes and weak at the fine motor work underneath.

Cost and speed are the catch

Simulation can pause while the model thinks. A factory line cannot. UBTech tested Astra on a real humanoid, and letting it control the arm directly took over an hour and cost about $579 without grabbing a Rubik's cube. When Astra only planned and a specialist action model did the grasping, the job finished in a few minutes at a small fraction of the cost.

Executives at several Chinese robot makers described similar problems: slow responses, stalls and very high per-call costs. Google's Gemini Robotics 2.5, tested in Mercedes-Benz plants, also uses two layers, one for tracking the workflow and one for controlling arms and hands.

Why Chinese firms keep investing

In the same week, the supply chain moved on several fronts. Examples include a compute pack for Unitree robots, a chip venture between UBTech and MetaX, a Beijing action plan for embodied intelligence, and a listing-preparation filing by a gripper and dexterous-hand supplier.

The bet is that whoever defines this middle layer avoids paying rent to AI labs above and chip vendors below. The risk is real: models improve fast. Anthropic's Fable 5.2 went from 27.6% to 64.9% on one simulated dual-arm test in a single generation, while GPT-6 Astra scored 29.8% on the same internal table. The article argues that lasting advantages will be physical, such as durable joints and real factory data.

What to watch

Try it

When you see a robot demo, ask three questions: was it real or simulated, how long did each action take, and what did each successful run cost?

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by tmtpost.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news