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
- Muse Charm is to phones what phones were to computers, argues Meta researcher Matt Deitke — alex_conneau · 2026-09-27
- Open-source VLA + RL pipeline ranks 1st of 62 teams folding garments on a two-armed robot — chris_j_paxton · 2026-09-27
- Insider confirms Muse Charm direction: hardware existed, OS3 completes the agentic stack — SimonBalmain · 2026-09-27
- AI Researcher Dietterich Questions Robotaxi Safety Culture: Too Slow to Fix Problem Behaviors — tdietterich · 2026-09-27
- E2E mobile manipulation on Agility Robotics Digit v4 works out of the box at IROS — chris_j_paxton · 2026-09-27
- Meta Connect 2026 hardware: new VR glasses, Ray-Ban Meta Gen 3 and Muse Charm — technextpreneur · 2026-09-27