Humanoids in factories: 15 weeks of training still needs intervention every 6 minutes
ZeeshanZiaML · x · 2026-11-11
VLA/manufacturing researcher Zeeshan Zia uses real data from a top-tier lab to challenge the humanoid-in-factory narrative.
- Task: picking crankshafts one-by-one from a tote onto a conveyor in a real factory — highly standardized conditions (same part number, bin size/color/placement), with a VLA fine-tuned via BC.
- Result: after 15 weeks of teleop intervention data, a human still must intervene roughly every 6 minutes.
- His verdict (7+ years training video activity models for manufacturing): this doesn't scale. Best case, companies repeat Agility's fate — pigeonholed into 4-5 narrow tasks per factory with broken sales cycles and unit economics; worst case, a 'call center' model fine-tuning per-robot weekly with one teleoperator per 5 robots at best.
- Extra arguments: real factories rarely have more than 5 people doing the same task simultaneously, teleoperators can't context-switch across physical tasks, and task diversity plus slow deployment velocity means post-training yields no cross-task generalization.
- The attached slide shows LBM 2.0's MTBI curve: 340.3s (98.2% autonomy), claimed to support 50 robots with 3 remote assistants.
More from Embodied
- Impossible Metals tests AI deep-sea mining robot at 4,000m, grabbing 450g of nodules — sven_ai · 2026-11-11
- Sharpa Humanoid Robot Takes Polaroid Photos at Shanghai CIIE — Scobleizer · 2026-11-11
- LeRobot 0.7 streams terabyte robotics datasets from the Hub, cutting data wait to 0.52% of step time — LysandreJik · 2026-11-11
- Fable 5.2 designs a Tamagotchi across KiCad, Inventor and Blender — ZeYanjie · 2026-11-11
- Compact exoskeleton glove with encoders and tactile sensing shown at CoRL 2026 — ZeYanjie · 2026-11-11
- Wheeled bases, not legs: why production-test humanoids should skip bipedal design — AIFlow_ML · 2026-11-11