Artificial Intelligence · Projected milestone
World models that train robots
90% by 2028 · including earlier years
The year is an editorial timing estimate. The percentage applies to occurrence by the stated deadline. These estimates are not calibrated odds.
AI trained on video simulates the physical world so robots can practise tasks in imagination before trying them.
Explore this milestone on the interactive mapWhat the forecast means
Milestone stage: deployment
demonstrated
World-model/action-model components used for physical robot learning; broad reliable robot operation remains separate.
Completion criteria: Two independently inspectable physical-robot studies show held-out task improvement against a non-world-model baseline, including failed trials, intervention counts and real execution. A visually convincing generated video alone is insufficient.
This page describes a forecast, rather than an achieved result. Read the methodology and assumptions.
Built on
Led to
- Humanoids on full factory shiftsForecast · ~2028influences
- Robots learn a job in a dayForecast · ~2029influences
- Lights-out car factoryForecast · ~2032influences
- Most robots designed by AIForecast · ~2038influences
- Whole worlds shared as one fileForecast · ~2046enables
- Robot-run industryForecast · ~2050influences
Sources & evidence
- Beyond VLAs: How world action models reshape robot manipulation 2026-08-04
NVIDIA describes video/world-model backbones and Cosmos 3 for robot action learning. This supports demonstrated world-model components, not a claim of universally accurate physics or reliable execution.
- Helix 2.5: Zero-shot generalization across 30 homes 2026-09-17
Figure reports three trained behaviors tested in thirty unseen homes without home-specific tuning, with aggregate whole-task success increasing from 9% to 56%. This is meaningful transfer and still falls short of dependable general household autonomy.
Record reviewed 2026-10-18