Ropedia has released Ropedia Xperience 2, a physical AI dataset that pairs metric LiDAR scans of real rooms, each with a 3D Gaussian splatting model, with head-mounted recordings of people performing tasks.
The Ropedia Xperience 2 paper, posted to arXiv on November 12 by a 21-author team, describes 1,589 released scene scans and a collection of 4,230 egocentric recordings covering 74 tasks and 131 hours. Capture took place in homes, restaurants, stores, offices and outdoor sites, with one operator and no robots or motion capture rigs on site.
Each scene is scanned once with a handheld LiDAR device carrying two fisheye and two wide-angle cameras. The pipeline exports posed fisheye images and a sparse model in COLMAP format, a triangle mesh, and a 3D Gaussian splatting model, all in one metric coordinate frame. The authors use the splat model to render each scene from arbitrary viewpoints, including the camera poses recorded by the wearer.
Task recordings come from HOMIE Gen 2, a head-mounted recorder with four global shutter fisheye cameras, an IMU and four-channel audio. An open-source tool, xperience-register, aligns each recording to its scan by matching frames against the scan images and mesh. Its documentation states that a point cloud or 3DGS model alone cannot replace those inputs.
The splats also appear in the robot learning demonstrations. Virtual teleoperation runs in a simulator scene built from a scan's mesh and 3DGS model, retargeted humanoid trajectories are replayed inside the Gaussian reconstruction, and converted simulator scenes are checked view by view against renders of the 3DGS.
The Hugging Face release currently holds a first batch of 126 of a planned 1,500+ scenes, about 1.07 TB. Each scene folder includes a gaussian_splat.ply with degree-zero spherical harmonics, an OBJ mesh, the COLMAP sparse model, and fisheye images with masks, alongside per-recording motion, pose and caption files. The data carries a CC BY-NC 4.0 license, and the registration code is Apache 2.0.
The first batch is available on Hugging Face.
