NTU team open-sources RopediaW0, an embodied world model trained on 10K hours of human egocentric video

青稞AI · wechat · 2026-11-05

A team from MMLab@NTU, Ropedia, Duke, UCSD and Tsinghua released RopediaW0, an open-source embodied world model that transfers across robot embodiments (humanoids, arms, grippers, dexterous hands) by unifying locomotion as camera poses and manipulation as projected skeletons in a shared image coordinate system.

Data: Xp-Data, built via a 5-stage pipeline from 14 public egocentric datasets plus robot data — 10K hours of human first-person video and 3K hours of robot video, with geometry/kinematics/semantic annotation layers.
Model: latent video diffusion in dense 5B and A14B MoE (27B total, 14B active, based on Wan2.2-I2V); pre-trained on human video, then mid-trained on mixed data.
Results: A14B hits 17.36dB zero-shot on Ψ0, beating fully-trained Cosmos3-Nano (16B, 16.55) and Cosmos3-Super (64B, 15.73). Human pretraining data shows clear scaling gains for both human and robot prediction. Distilled 4-step students match 50-step teachers (7x faster), enabling minute-scale rollouts and real-time Quest3 VR world-model interaction. On LIBERO-Plus with only 3 demos: 23.97% vs FastWAM's 21.44%.

Paper: arXiv:2610.26855; code open-sourced on GitHub.

Original post →

NTU team open-sources RopediaW0, an embodied world model trained on 10K hours of human egocentric video

More from Embodied