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.
More from Embodied
- Lighthouse extends Obsidian welding foundation model to run on any robotic platform — bennash · 2026-11-11
- Sipeed's MaixStation Dock PCB returns, 14-layer SOM still in fab — AIFlow_ML · 2026-11-11
- NAE EU-US Frontiers of Engineering symposium explores AI-driven materials, robotics, neuromorphic computing — ProfBuehlerMIT · 2026-11-11
- Paper link for FLEX: robots generalize language corrections into preferences — JafarNajafov · 2026-11-11
- MIT and NYU's FLEX lets robots generalize a single spoken correction into broad preferences — JafarNajafov · 2026-11-11
- AI lets robotics research thrive in labs with no robotics experts — 4310sy · 2026-11-11