DeFM: ETH team builds a DINO-style foundation model for depth images in robotics
ChongZzZhang · x · 2026-10-28
Manthan Patel and colleagues at ETH Zurich's Robotic Systems Lab released DeFM, a self-supervised foundation model for depth images, accepted to IEEE T-RO. Trained with a DINO-style distillation objective on 60.4M depth images from 18 real/synthetic/MDE-derived datasets, it keeps metric awareness across scales via a novel input normalization. One frozen backbone works across navigation, manipulation and locomotion; 12 ready-to-use models (3M–304M params) hit SOTA on depth-based benchmarks with strong sim-to-real transfer. Weights, code and demo are out.
More from Embodied
- Machenta CEO bets on robots that learn tasks from a single demo — Ed__Johns · 2026-10-28
- Microduck team spends a month in Shenzhen to finalize mass production of its humanoid robot — victormustar · 2026-10-28
- crowdbricks: open-source batching hits 10k motion generations/sec on one RTX 4090, 26x faster — TinfoilTricorn · 2026-10-28
- Tesla Optimus lines target 1M robots/year, public sales eyed for late 2027 — anthara_ai · 2026-10-28
- Dream-1 robot arms grab limes zero-shot from a single prompt in restaurant kitchen demo — ___Mufasaa · 2026-10-28
- Metatron: a System 1 controller that lets coding agents command humanoid robots — ChongZzZhang · 2026-10-28