Shangpin Home Collection Open-Sources WorldSimReady-Home Dataset for Robot Training

Getting humanoid robots into homes is less a hardware problem than an environment problem. Floor plans vary widely, furniture layouts differ, and daily routines are complex. Robots that shine in labs often struggle in real homes—a challenge known as the Sim2Real gap. The gap is not just technical; it reflects the enormous variability of real domestic spaces.

Closing that gap requires repeated training in near-real home environments, but real-world data collection is expensive and scenes are hard to reproduce. WorldSimReady-Home, an open-source simulation dataset released by Shangpin Home Collection and Tangyuan Technology, targets this challenge. It aims to give robots a controllable, scalable training ground before they enter real homes.

WorldSimReady-Home's first open-source release includes 100,000 square meters of high-fidelity home scenes, 10,000 interactive assets, and 1,000 standardized robot simulation task examples. It is not a static asset library. Using the initial scenes as seeds, the dataset can generalize across spatial layouts, object arrangements, materials, and task conditions to generate tens of thousands of differentiated home simulation scenes, suitable for training and testing home robots of different forms and functions.

Shangpin contributed more than 3 million real floor plans and over 30 million design schemes, providing real home spaces and high-quality 3D design assets. In terms of quality, scene geometry reaches millimeter-level precision, semantic and material accuracy exceeds 99%, and key physical property accuracy exceeds 95%. Each asset carries four layers of core information—geometric accuracy, semantic completeness, physical plausibility, and interaction usability. Besides RGB images, depth maps, and semantic segmentation, the dataset provides physical and interaction properties such as object joints, mass, friction, damping, and collision.

Traditional real-world data often fits only a specific robot body. WorldSimReady-Home decouples the environment from the robot body at the engineering level and supports quick integration of standard models such as URDF. Humanoid, quadruped, and wheeled robots can reuse the same scene and task data. This means one set of home scene data can serve different robot forms and functions. Developers can directly use the dataset for simulation training in navigation, object manipulation, and complex long-horizon household tasks.

Compared with real-robot debugging, simulation training offers lower trial-and-error costs, controllable experimental conditions, and large-scale task reproduction. Much of the training and evaluation can be completed in simulation, reducing reliance on physical trials. In plain terms, it builds a digital home training ground where robots can repeatedly practice navigation, grasping, placement, and long-horizon chores without risking damaged furniture or accidents—then transfer learned policies to physical robots.