Ropedia Xperience 2 Opens 1,589 Scans and 131 Hours of Task Video

Ropedia Xperience 2 bundles 1,589 metric-scale 3D scans with 4,230 egocentric recordings covering 74 task types across 131 hours, collected at 596 locations with a handheld LiDAR scanner and a head-mounted multi-camera rig called HOMIE. The release leans residential: 1,389 of the scans, 87.4% of the dataset, come from home interiors.

The authors frame the release against today's robot-control models. "Large foundation models such as GPT-6 Astra and Claude Fable 5.1 have recently been shown to control robots directly from visual observations," the paper reads. "They nevertheless remain brittle on long-horizon, contact-rich tasks in real-world environments." The data strategy follows from that: "Capture thus scales with operators rather than robots, and the captured data can be expanded into robot data without further on-site collection."

Three benchmarks ride on the scans. On a 500-question spatial-intelligence slice, Gemini-2.5-Pro leads at 74.20% overall and GPT-6 Astra High trails at 71.80%, but no model exceeds 34% on distance or 55% on translation. For 3D object detection on the same scenes, the paper is blunt: "Boxer leads on all three metrics. The other methods stay below 7 AP₂₅, and LLaVA-3D and SceneScript produce no box that reaches IoU 0.25." No method clears 7 AP₅₀ at all.

The robot-learning demonstration converts captured human motion to a wheeled humanoid. Median hand-tracking error lands between 0.40 and 0.55 cm, and the trajectory satisfies physical constraints in 100% of frames. An eight-subtask kitchen-to-bedroom routine (pick up mug, open microwave, insert, close, remove, close, open bedroom door, place mug on bedside table) completes 30 of 30 simulated rollouts without task violations.

It lands into a crowded space: 205 robotics stories have crossed our tracker in the last 90 days, including Runway's Praxis-1 open release yesterday.