Read original
arXivHyesung LeePapers88

RORA: Realistic Object Reconstruction with Articulation

RORA presents an end-to-end pipeline for reconstructing simulation-ready articulated objects from a single static object video, avoiding the need to scan dynamic motion. It combines 3D Gaussian Splatting for photorealistic appearance with mesh geometry for physical interaction. The pipeline performs convex decomposition, human-guided part grouping, Gaussian-to-mesh binding, and automatic joint-axis suggestion based on local boundary geometry. The authors report precise articulation results on PartNet-Mobility-v0 and real objects, and demonstrate reconstructed assets in Unreal Engine and NVIDIA Isaac Sim for real-time dexterous-hand manipulation tasks.

Why it's worth reading

As robot simulators increasingly require physically interactive assets, RORA offers a concrete route from one static video to articulated objects while exposing the remaining human-in-the-loop and evaluation constraints.

Tags

机器人仿真3DGS物体重建关节建模sim-to-real灵巧手Isaac Sim