论文概要 (Paper Overview)
Field: Computer Vision Authors: Tianqi Liu, Hanfei Zhao, Qifeng Liang, Hui En Pang, Jiarui Xu, Sayang Mu, Penghao Qiu, YingSheng Zhang, Jun Wang, Ziang Cao, Mingze Sun, Yukang Cao, Fangzhou Hong, Zhaoxi Chen, Ziwei Liu arXiv: 2610.26855
Summary
World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. The authors present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments.
Key Contributions
- Xp-Data dataset: 10k hours of egocentric human data plus 3k hours of robot data, with unified multimodal annotations and systematic quality control.
- Unified kinematic representation: locomotion is encoded through camera poses, while manipulation is encoded through projected skeletons. Robot skeletons combine embodiment-specific forward kinematics with extensive manual keypoint annotation and correction.
- State-of-the-art results: best performance across five cross-embodiment benchmarks — DreamDojo, Psi-0 (Ψ₀), omega-HOME (ω-HOME), Xp-Teleop, and AgiBot.
- Scaling insight: scaling human pretraining data improves performance on both human and robot evaluations.
- Real-time deployment: after distillation, the model supports real-time minute-level rollouts, enabling live VR interaction, with practical applications in policy evaluation and model-based planning.
- Paper: https://arxiv.org/abs/2610.26855