A real-world test of OpenAI’s GPT-6 Astra has put a hard number on the gap between a robot that understands the world and one that can feel it. On a task that asked a robotic arm to pick a block and drop it into a bowl, Astra reached a 95 per cent success rate, far above the 40 per cent posted by Claude Fable 5.1. But when the same model was asked to align and insert a puzzle piece into a millimetre-precision slot, the success rate collapsed to 10 per cent. Around the same time, Sam Altman said on a podcast that OpenAI would build its own humanoid hardware.
The gap is not where the industry thought it was
For the past year the consensus held that the bottleneck in embodied AI was cognition: robots could not understand instructions or read a scene. So the field piled into vision-language-action (VLA) models and world models to make robots “smarter”. Astra’s 95 per cent proves a large model can indeed supply that missing cognition. With no task-specific training, it drives a robotic arm to a 95 per cent grasp rate. The smoothness of that transfer has alarmed founders building bespoke embodied models; one former embodied lead at D-Robotics argued openly that VLA and world-model approaches may no longer be worth the investment.
The 10 per cent assembly rate exposes a different dimension. Dropping a block into a bowl is macro spatial operation. The robot only needs to know where the block is, where the bowl is, and how to move between them. That is exactly what large models are good at: spatial understanding and semantic generalisation. A team at Tsinghua’s Institute for Interdisciplinary Information Sciences concluded after testing that Astra generalises well in semantics and space but is weak in physical generalisation.
Inserting a puzzle piece demands micro force-feedback control. The robot must sense tiny changes in resistance at its fingertips, judge whether the piece is tilted, and adjust force and angle in real time, all inside a millimetre of space and a few milliseconds of time. A camera cannot see it, a language model cannot reason about it, only a tactile sensor can capture it. As one analyst put it: Astra can understand the world but cannot touch it.
Capital is already voting with real money
That 10 per cent figure is the strongest endorsement the tactile-sensing track has ever received. For two years, tactile startups pitched a story that tactile sensing matters for fine manipulation, but investors struggled to price how much that matters. Astra gave them a benchmark: without touch, fine-assembly success is capped at 10 per cent, and the road from 10 per cent to 90 per cent runs through sensor precision, data scale, and force-control maturity.
The primary market has already moved. In the first seven months of this year new funding in tactile sensing passed 7 billion yuan, and at least two companies reached unicorn status. Paxini alone has raised 3.5 billion yuan, a global record for the field, at a valuation above 10 billion yuan. On the secondary market the signal is sharper still: when Fulaixin signed a 100,000-unit tactile-sensor order, its shares hit limit-up the same day. Hanwei’s subsidiary Nengsida builds 20 million flexible tactile sensors a year at a gross margin above 40 per cent, and Keli shipped more than 1,500 force sensors in 2026 with monthly output past 1,000 units.
The real turning point is the price
The two markets price on different logic. The primary market buys the thesis that touch is a necessity, anchoring to a trillion-yuanlong-term penetration. The secondary market buys whether orders land. Both eventually converge on one question: can tactile sensors be mass-produced, and can the price fall? Paxini cut the cost of a high-end tactile sensor from the hundred-thousand-yuan tier to the hundred-yuan tier, a hundredfold drop that turns a premium option into a mass-production standard. That is the moment the track moves from story to business.
Altman says OpenAI must build humanoids. But Astra’s 10 per cent already tells the industry that large models solve the brain and not the touch. That last centimetre of tactile sensing is the segment of China’s humanoid supply chain that is running fastest.
Editor’s note: This is an adapted translation of the original OFweek Robotics report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-09/ART-8321205-8110-30703910.html.