Humanoid in factory: still needs a human takeover every 6 minutes after 15 weeks of training
ZeeshanZiaML · x · 2026-11-11
Robotics practitioner Zeeshan Zia argues humanoids don't scale in manufacturing, citing data from a top-tier lab: on a highly standardized crankshaft pick-and-place task, a fine-tuned VLA (behavior cloning) still needed teleoperator intervention roughly every 6 minutes after 15 weeks of teleop data — 340.3s MTBI at 98.2% autonomy.
Key points:
- Economics don't work: real factories rarely have more than 5 people doing the same task simultaneously, and teleoperators can't context-switch across physical tasks; 1 operator per 5 robots is optimistic.
- No cross-task generalization: task diversity and slow deployment velocity mean post-training won't generalize beyond commodity mid-training data.
- Two outcomes: best case, companies get pigeon-holed into 4-5 narrow tasks per factory like Agility, with broken unit economics; worst case they become a "call center" — weekly per-model fine-tuning plus teleoperators.
The author, with 7+ years training video activity models for manufacturing, concludes the approach "just doesn't scale."
More from Embodied
- GEN-1.5 robot assembles engine housings 100 times in a row with Foundry Robotics — ishabytes · 2026-11-11
- Laser-as-a-sensor: monolithic VCSEL oxygen detector shrinks sampling volume to subnanoliter — jwt0625 · 2026-11-11
- Beni robot, #1 personal robot on Kickstarter, officially launches with self-recovery and 4K60 — Scobleizer · 2026-11-11
- Berkeley's Social Humanoid teaches Unitree G1 socially responsive interaction — berkeley_ai · 2026-11-11
- Robotics is not solved: frontier models are ~6 orders of magnitude off real assembly lines — ZeeshanZiaML · 2026-11-11
- MacBooks do have a hinge angle sensor—verified with Codex, prompt included — lipeng0820 · 2026-11-11