Research

World models take a leap, and robot training may never be the same

Learned simulators now predict physical interactions accurately enough to train robot policies almost entirely in imagination, slashing real-world training time.

By James Park, Robotics Correspondent — TOKYO

TOKYO — Learned world models — neural networks that predict how physical scenes evolve — have crossed a fidelity threshold with outsized consequences: robot skills trained almost entirely inside them now transfer to real hardware with only minutes of physical fine-tuning, according to results presented by three separate groups this month.

The convergence matters because physical trial-and-error has been robotics' binding constraint. A policy that once needed weeks of supervised hardware time can now rehearse millions of variations overnight in imagination.

The groups all trained on large corpora of real interaction video, and all report the same open problem: rare, high-stakes physics — sloshing liquids, deformable objects, humans behaving unpredictably — still demands real-world data.

Enable JavaScript to read the full story on Neural Daily News.