Figure's Helix 2.5 still fails ~44% of trained tasks — and '30 unseen homes' are showroom-clean

YakFull8300 · reddit · 2026-09-19

A Reddit critique takes apart Figure's Helix 2.5 announcement of "zero-shot generalization to 30 unseen Bay Area homes":

• The robot still fails 44% of full-task trials on the very three tasks it was specifically trained for, yet coverage jumps straight to "general-purpose home robot";
• Demos have been the same loop for years — dishwasher, laundry, towels, toys, bed-making — with no truly untrained household task ever shown;
• The "30 unseen homes" are architecturally monotonous and showroom-clean: clear floors, unobstructed walkways, objects conveniently placed at task sites;
• Brett Adcock's "it's never seen this pillow before" moment gets mocked: it's a generic white pillow.

Original post →

More from Fun

• Dev spends two weeks having Astra 3D-model every object in a three.js scene — danbri · 2026-09-19
• Speech-to-text maxing: the new way to prove your posts weren't written by Claude — willcb · 2026-09-19
• An iPhone app that only shows yesterday's weather is winning over users — genmon · 2026-09-19
• Meme: 'Frontier Labs Marketing 101' Mocks AI Lab Hype Playbooks — AlexTensor · 2026-09-19
• Siri refuses to define a word while driving — and still refuses after parking — tomchapin · 2026-09-19
• Civilization II Lord of the Rings Scenario Lets Frodo Carry the Ring via Aircraft Carrier Flag — xuenay · 2026-09-19

Fun channel →