Stanford/PKU paper: never train inside the world model — QWM uses it only for action selection

rohanpaul_ai · x · 2026-08-23

A new Stanford and Peking University paper, "Q-Learning With World Models" (QWM), argues that training a robot policy inside a learned world model means the policy inherits every mistake the model makes — and those errors pile up as tasks get longer and images get messier. QWM never trains anything inside the model: the world model doesn't touch training at all, it only helps the robot choose actions. Paper: arxiv.org/abs/2608.17163.

Original post →

More from Embodied

Embodied channel →