DeFM Depth-First Foundation Model: Why Anchor to Depth, Not Estimator Features
chris_j_paxton · x · 2026-10-28
Chris Paxton highlights DeFM, a depth-first foundation model for robotics, and a discussion of a key design choice: why use a separate encoding step after depth extraction instead of pulling intermediate features from the depth estimator itself?
- Monocular metric depth estimation is far from solved; locking into one model's intermediate features would freeze the pipeline and force retraining to migrate to better models
- Anchoring to the depth output acts as an output standard that future depth models can be expected to follow
- Metric depth matters beyond robotics too (e.g., autonomous driving), making the standard durable
More from Embodied
- A 1-inch robot roach that crawls into pipes and smells rust — Scobleizer · 2026-10-28
- Ex-Apple team launches Silica: joint actuators 50x more sensitive, pre-orders open — ycombinator · 2026-10-28
- edwinarbus refuses to wear AI glasses on the train: "Don't look at me" — edwinarbus · 2026-10-28
- Tesla Is Grinding Through the Same Struggles Waymo Endured, Analyst Notes — binarybits · 2026-10-28
- Kodiak signs first-of-its-kind deal to deploy its driver in 75 driverless DTL trucks — Scobleizer · 2026-10-28
- SnowPod desktop metal 3D printer debuts at TCT Shenzhen, prints stainless steel — TinfoilTricorn · 2026-10-28