Foundation models

A foundation model is a large network pretrained once on massive, diverse data, then reused across many downstream tasks — either frozen as a feature extractor, lightly fine-tuned, or prompted. The term entered SLAM vocabulary because a wave of such models turned out to be directly useful as perception backbones: CLIP (image-text embeddings), SAM (promptable segmentation), DINOv2 (self-supervised visual features), Depth Anything (monocular depth), and the DUSt3R family (two-view 3D reconstruction from raw image pairs).

What changed

Why this matters is best seen against the previous norm. Classical learned modules (SuperPoint, MonoDepth) were trained per-task on modest datasets and inherited their biases; deploying them in a new domain often meant retraining. Foundation models change the economics:

The families and what each provides

FamilyTrained on / howWhat SLAM gets from it
CLIP / SigLIPimage-text pairs, contrastiveopen-vocabulary semantics for maps and place descriptions
SAM / SAM 2promptable segmentation at scaleclass-agnostic instance masks for object-level and dynamic SLAM
DINOv2self-supervised ViTrobust dense features for matching, VPR, and correspondence
Depth Anything (V1/V2), Metric3Dlarge-scale depth (self-)supervisiondense monocular depth priors where no depth sensor exists
DUSt3R / MASt3R / VGGTposed image pairs/collectionsfeed-forward pointmaps, matches, and poses — geometry as inference

Knowing this menu tells you what you can now get “for free” when designing a system — and which of the classical Level 3/4 modules each model could replace.

How SLAM systems actually consume them

Trade-offs to respect

The practical recipe

The pattern emerging across Level 5 and beyond: use foundation-model features for the robustness-critical, semantic, or ill-posed parts of the pipeline (place recognition, wide-baseline matching, monocular depth, open-vocabulary labels), and keep classical estimation for the precision-critical parts (pose optimization, mapping consistency). The most recent systems — MASt3R-SLAM, ConceptGraphs-style semantic mapping — are exactly this: thin geometric layers over large pretrained models.

Why it matters for SLAM

Foundation models are reshaping what a SLAM front-end is: instead of training task-specific networks, modern systems increasingly compose frozen pretrained backbones — and the most recent SLAM systems (MASt3R-SLAM, ConceptGraphs-style semantic mapping) are thin geometric layers over such models. Knowing what each family provides (CLIP: semantics, SAM: masks, DINOv2: robust features, Depth Anything: depth, DUSt3R: geometry) tells you what you can now get “for free” when designing a system.