01
until last week
the only way to know if a robot model worked
was to run it on a real robot
02
weeks per test
one model at a time
hardware that costs more than a car
03
yunfei xie and four colleagues submitted a paper on october 22nd
proposing a cheap calculation you run before touching any hardware at all
04
the idea traces back to plato
and to mit researchers who showed in 2024
that capable systems converge on the same picture of reality
05
a lab with one robot
can now screen hundreds of candidates
06
instead of ten