Expert verdict on GPT-6 Astra as a robot brain: great on benchmarks, can't fold laundry — general-purpose robots not until early 2028
From the Robots digest for 2026-10-20
What happened: Also published early on October 20, 2026, a writeup of commentary by robotics researcher Xu Huazhe following OpenAI's release of GPT-6 "Astra": three experts delivered a split verdict, crediting the model with breaking robot benchmarks — while noting it is merely incremental on coding and agent tasks — and pointing out that benchmark performance does not translate into basic manipulation. The takeaway headline conclusion: Astra still cannot fold laundry at all, and general-purpose robots are unlikely to arrive until early 2028. This is commentary and projection, not a capability announcement; nothing new shipped.
What's genuinely new: A concrete, dated timeline from a named voice in the humanoid reality-check storyline — the same skeptical strand that has carried DeepMind's "GPT-2 stage" caution from Keerthana Gopalakrishnan, the IFR's Jane Heffner, and Mark Cuban's dismissals. The "GPT-2 stage" framing said reliability, not raw capability, is the binding constraint; this piece supplies the same argument from the benchmark direction: model scores rising while household-grade manipulation sits at zero. A timeline of early 2028 is more specific than the thread has carried before.
Watch: Whether other experts align on early 2028 or push the date further; whether the next frontier-model release gets tested for the laundry-grade manipulation gap rather than benchmark gains; and how this verdict is received against vendor timelines already on file, several of which aim earlier.
Links: BigGo Finance — Xu Huazhe: Astra breaks robot benchmarks, but can't fold laundry at all