10M‑shot robotics dataset
Ropedia released Xperience‑10M: a 10 million egocentric interactions dataset with ~10k hours of multimodal data (video, audio, depth, mocap) aimed at embodied AI and world models — a major training resource for manipulation and real‑world robot learning. Large egocentric corpora like this are what foundation models for robotics will need to scale. (x.com)
Ropedia published the dataset release on March 16, 2026 and is hosting dataset assets and documentation through its site and partner repositories. (ropedia.com) The public dataset includes synchronized multi‑stream recordings such as six fisheye camera feeds, stereo depth, audio, camera pose, hand motion capture, full‑body mocap, IMU telemetry and hierarchical language annotations. (huggingface.co) A downloadable sample package with raw videos and a 3D/4D visualization (.rrd) is published on Hugging Face and contains example episodes (e.g., a labeled coffee‑making clip) plus visualization assets. (huggingface.co) Ropedia released an open-source HOMIE‑toolkit on GitHub with data loaders, visualization helpers and example scripts under an MIT‑style repo to help researchers load annotations and render depth, skeletons and point clouds. (github.com) The company was incorporated in Singapore in 2025 and lists Zhaoxi Chen (CEO), Fangzhou Hong (CTO) and Prof. Ziwei Liu (Chief Scientist) among its founding team; media reports describe a seed round in the "tens of millions of dollars" from US and Asian investors. (databasesets.com) Hugging Face’s dataset card enumerates 19 supported tasks (video classification, image‑to‑text, depth estimation among them) and Ropedia’s release notes highlight the dataset is intended for foundation‑model pretraining and downstream embodied‑AI evaluation. (huggingface.co)
