Figure's 237 of 420 success rate turns into a fight
Tony Zhao says failing half the time is not useful work, and two other researchers say publishing the number at all deserves credit.
Figure's four hours of robot footage from 30 rental homes ended the week being argued over a single number pulled from Figure's own blog post.
Brett Adcock posted the video, saying the robot did zero-shot work in 30 Bay Area rental homes, meaning it entered places it had never seen and started working without any new training. That is the earlier part of the story. What is new is the pushback.
Tony Zhao quoted the blog's 237 out of 420 success rate and said that is not the same as doing real useful work. His definition is two things together, generalization plus reliability, and he argued Figure has shown the first and not the second. He opened with "Love my friends at figure", so this is a disagreement inside the field rather than a dunk from outside it.
Two researchers defend the number
Keerthana Gopalakrishnan pushed back on the pushback. Her argument is that plenty of companies overclaim and stage demos, so a lab that publishes its real hit rate should be rewarded for it, not criticized.
Chris Paxton took the same side. He said showing real success rates, even when they are not high, shows confidence that the numbers will improve.
What is actually established
The 237 out of 420 figure comes from Figure's own write-up. Nobody outside Figure has independently reproduced it, and none of the sources here claim to have.
Adcock's framing is that generalization is the hard part. In his words the holy grail for robotics is doing work in unseen places, and Helix 2.5 was built to test whether a humanoid can enter a home it has never seen and immediately get to work with its whole body, on its own.
So the split is not really about whether Figure is telling the truth. All three researchers appear to accept the number. The argument is about what counts as a milestone. Zhao's position is that a rate near half is a research result, not a working product. Gopalakrishnan and Paxton's position is that the alternative to publishing a middling rate is publishing nothing, and the field has enough of that already.
Let’s not shame people for being transparent about their success rates.
failing half the time is not “doing real useful work.
I think its great that theyre showing real success rates, even if theyre not super high. Shows confidence in improvement