SILICA: diffusion priors crack glass segmentation and depth sensing (IROS 2026)
rsasaki0109 · x · 2026-08-25
SILICA, a paper from IIIT Hyderabad and collaborators accepted at IROS 2026, repurposes text-to-image diffusion priors to jointly predict glass segmentation and glass-aware monocular depth.
- Motivation: LiDAR and RGB-D sensors largely fail on transparent surfaces like glass doors, corrupting maps and risking unsafe navigation
- Method: CLIP-conditioned diffusion priors let the two tasks learn together, requiring no paired real-world glass depth annotations; predictions fuse with conventional sensors to recover accurate metric depth on glass
- Data: introduces Mirage 18k, 18k+ manually annotated images with segmentation labels and a glass-aware depth benchmark
- Results: strong zero-shot generalization; deployed on an autonomous wheelchair and Husky platform, achieving collision-free navigation past glass walls and doors
Paper, code, and dataset are all public.
More from Embodied
- Apple unveils new Mac Studio with M5 Ultra, up to 512GB unified memory — themixtergames · 2026-08-25
- Robotics Outlook: Intelligence Won't Be the Bottleneck Next Year, Supply Will — Rewkang · 2026-08-25
- Matic robot fails: Bread tie stuck in vacuum duct — tristanbob · 2026-08-25
- Humanoid robot invents new running style to avoid shoulder overheating — CSProfKGD · 2026-08-25
- China unveils voice-summoned self-driving robot toilet — Polymarket · 2026-08-25
- NTU's World-Model-Centric Racing Agent Achieves 88.3% Interaction Success — jiqizhixin · 2026-08-25