Skip to content
RobothecaBeta
Log in

DeFM (depth foundation model)

  • AI model
  • Perception

DeFM is a vision backbone pretrained on 60M depth images via self-distillation, meant for robotic perception and sim-to-real transfer. It is a DINO-style encoder for depth inputs.

Made by Robotic Systems Lab, ETH Zürich, Switzerland (checked on github.com/leggedrobotics (opens in a new tab)).

Source on GitHub: github.com/leggedrobotics/defm (opens in a new tab). Facts read at commit 63ec5e1 (opens in a new tab).

Licenses

Code
Apache-2.0 (opens in a new tab)

Read by Robotheca from LICENSE (opens in a new tab) (Apache-2.0) at commit 63ec5e1 (opens in a new tab), on 16 Oct 2026.

Commercial use
Yes

What it does

  • Pretrained on a curated set of 60 million depth images using self-distillation.
  • Produces semantically rich features learned from depth images alone.
  • Uses a three-channel metric-aware normalization that preserves metric depth across scales.
  • Distills the largest ViT-L model into smaller CNNs, the smallest with about 3M parameters, for robot policy learning.
  • Provides a preprocessing function that converts depth maps in meters into the model's three-channel input format.
  • Loadable via TorchHub, with pretrained checkpoints for ViT, ResNet, EfficientNet and RegNetY families on Hugging Face.

Written by Robotheca from the project's README (opens in a new tab) at commit 63ec5e1.

Getting started

Difficulty
Intermediate

Maintenance

Status
Maintained: a commit in the last 12 months
Newest commit
27 Apr 2026 on main
Latest release
None published on GitHub

Checked on GitHub on 16 Oct 2026.

Links

Robots that can run it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.