The Secret Wasn't a Bigger Model: 500 Randomized Demos Lift Dream-1 Score 159%

paigeinsf · x · 2026-11-19

The gain came not from a bigger model but from 500 demonstrations, each starting from a randomly scrambled layout, lifting Dream-1's score 159% from 8.5 to 22.

Before fine-tuning, Dream-1 barely reached past the middle of the table — 4 of 16 toy peppers near the edges; after, 11 of 14. It even learned a rule it used to ignore: lining up cubes using only a ruler, going from a zero-shot score of 1 to 41 after fine-tuning.

Original post →

More from Embodied

Embodied channel →