Cloth is the material that most reliably gives away a render. It is anisotropic, it is retroreflective at grazing angles, its highlights are governed by thread direction rather than surface normal, and its appearance changes with both view and light in ways that a single analytic BRDF cannot hold. The standard response is to measure it, and measured-reflectance datasets for fabric have historically been small, because measuring one material well takes a rig and a day.
RoboCloth is 500 of them.
The paper — from Zhen Li (Simon Fraser University), Haoran Qin (Westlake University), Francesco Di Sario (University of Torino and Simon Fraser University), Daniel Rebain (University of British Columbia and Wayve) and Andrea Tagliasacchi (Simon Fraser University, Wayve and University of Toronto) — was posted on 15 October.
The acquisition is the paper
Each of the 500 materials is imaged in approximately 580 high-dynamic-range frames spanning a wide view-light hemisphere, with both the camera and the lighting robotically controlled during capture.
That phrase is doing a lot of work. A conventional gonioreflectometer moves a sensor and a source on calibrated arms and is accurate, slow and bespoke. A robot-mounted camera plus a robot-mounted light is less exotic hardware and far more throughput — but you then have to recover where everything actually was. The authors do that by refining camera poses with structure-from-motion and aligning them to the robot frame, which yields two things per material: sparse surface geometry, and structured per-point reflectance observations already in the shape a neural material wants.
The dataset ships with the calibration files, the observation tensors, and the train/test splits. That last item is the difference between a dataset and a release.
The two-stage pipeline, and why stage 1 matters
The authors demonstrate the data’s value with a deliberately simple reconstruction method:
- Stage 1 trains a shared MLP decoder across the whole corpus, learning cloth material priors.
- Stage 2 freezes that decoder and fits a dense 2048 × 2048 latent texture to each target material.
The split is the point. A per-material network has to learn, from one material’s 580 frames, everything about how fabric scatters light. A shared decoder learns that once from 500 materials, and the per-material fit then only has to say which fabric this is, spatially. The 2048² latent texture is what carries weave, print and local variation; the decoder carries “this is cloth.”
It generalises off its own rig, which is the hard part
Measured-reflectance work has a chronic problem: a model trained on one acquisition system quietly learns that system. The authors test for it. They evaluate on held-out materials from their own dataset and on the independent UBO2014 benchmark, and report that a decoder trained on their captures outperforms decoders trained on existing measured-reflectance datasets and analytic PBR fitting, transferring to unseen acquisition systems without modification.
Transferring to a different rig without modification is the claim that matters for anyone who wants to use this. It means the learned prior is about fabric rather than about one lab’s lights.
What you can actually do with it
The dataset is on the Hugging Face Hub under CC BY 4.0 — permissive, attribution-only, commercial use included. It was created on 14 October and had 1,656 downloads as of this writing, with no likes, which is the usual signature of a resource being pulled by pipelines rather than browsed.
Realistic uses today:
- Material priors for garment reconstruction. Anyone fitting appearance to clothing from photographs now has a cloth-specific prior that is not a repurposed object-scan prior.
- Ground truth for appearance capture research. 500 materials with calibrated view-light coverage and published splits is a benchmark, not just training data.
- A sanity check on analytic fits. The paper’s comparison against analytic PBR fitting gives a measured bound on how much a parametric cloth shader is leaving on the table.
What it is not: a drop-in asset library. These are reflectance observations and latent textures, not game-ready PBR sets. Converting one to a four-map stack is a lossy operation, and the whole argument of the work is that the loss is where fabric lives.
Sources
- RoboCloth: A Large-Scale Real Cloth Material Dataset for Neural Reflectance Reconstruction — arXiv:2610.21090
- Project page
- theialab/robocloth — code
- theialab/RoboCloth — dataset on Hugging Face