RoboCloth: 500 real cloths, 80B observations to replace 2003-era BRDF data
taiyasaki · x · 2026-10-20
Researchers released RoboCloth, a large-scale real-world cloth material dataset for neural reflectance reconstruction — ending the field's reliance on 100 BRDFs measured at MERL in 2003.
- 500 real cloth materials, 580 HDR frames each (≈80 billion observations total), captured with robotically controlled camera and lighting; poses refined via SfM and aligned to the robot frame
- Provides per-material sparse geometry and structured per-point reflectance observations ready for neural material training
- A two-stage pipeline (shared MLP decoder for cloth priors, then a frozen decoder fitted with a dense 2048×2048 latent texture) outperforms decoders trained on existing measured datasets and analytic PBR fitting, transferring to unseen capture systems without modification
- Dataset, calibration files, observation tensors, and splits released under CC BY 4.0 on Hugging Face
Team spans SFU, Westlake University, UBC, and Wayve Labs; paper and code are public.
Related event: RoboCloth: Massive Real-Cloth Dataset With 80 Billion Observations Released(4 posts)→
More from Research
- MITRA-lexicon web app: LLM summaries with linked evidence for Sanskrit entries — SebastianNehrd2 · 2026-10-20
- MITRA-lexicon: 2.18M-entry Sanskrit-English-Tibetan-Chinese lexica built with LLM alignment — SebastianNehrd2 · 2026-10-20
- Anna Ciaunica argues consciousness is a journey, not a thing — and why AI cannot have it — AnnaCiaunica · 2026-10-20
- Ablation: even/odd tokens, not 'cat', control the zigzag — and it fits only one probe — Sauers_ · 2026-10-20
- Researcher trains Qwen3-1.7B to obey activation-space instructions, drawing zigzag cat representations — Sauers_ · 2026-10-20
- How the zigzag LoRA was trained: per-layer gradient signals with output forced to match the original model — Sauers_ · 2026-10-20