DeMiAn: dense language annotations boost robot policy learning, cut compute 62%

rajammanabrolu · x · 2026-09-27

The pearls-lab team released DeMiAn (Dense Multi-aspect Annotation), a scaling lever for robot policy learning that uses a Qwen3.5-2B instructor to auto-generate dense language annotations from pixels — no new demonstrations needed. It gains +5 pts on RoboCasa365 and +9 pts on unseen composite tasks while saving 62% compute (1.3×10²⁰ FLOPs) at 1M-clip scale. Instantiated as DeMiAn-VLA (openpi/Pi0.5) and DeMiAn-WAM (Cosmos-Predict 2.5 + action head); code and dataset are on GitHub and Hugging Face.

Related event: DeMiAn Open-Sourced: Dense Language Annotations Cut Robot Learning Compute 62%(2 posts)→

Original post →

More from Embodied

Embodied channel →