Ropedia releases W0, a cross-embodiment world model with NTU, Duke, UCSD and Tsinghua
Ropedia released W0, a cross-embodiment world model built with NTU, Duke, UCSD and Tsinghua researchers.
Singapore-based embodied AI startup Ropedia has released W0, a cross-embodiment world model for mobile manipulation developed with researchers from Nanyang Technological University, Duke University, the University of California San Diego and Tsinghua University.
W0 takes initial frames, text descriptions, camera trajectories and hand actions as inputs and generates first-person future video — predicting how views and environments change during movement and how hand-object states change during manipulation. A unified action representation lets one model cover humans, robotic arms and humanoid robots, and the company demonstrated continuous 30-second-plus predictions of tasks including opening a refrigerator, collecting garbage, sorting fruit and making a bed.
The model ships in two sizes: a 5-billion-parameter dense model (W0 5B) and a mixture-of-experts model with 14 billion activated parameters (W0 A14B). Training ran in stages on roughly 13,000 hours of multimodal data — about 10,000 hours of first-person human video and 3,000 hours of real robot data — after pretraining on large-scale human egocentric video.
The company says W0 ranks first across five evaluation benchmarks and 11 tests, including beating NVIDIA’s Cosmos3-Super trained on full data without post-training — performance claims that have not yet been independently validated. It also reports using W0 as a simulated environment for policy evaluation: predicted task progress correlated with actual execution at 0.984 in simulation and 0.996 on real robots, and initializing the FastWAM world-action model from W0 5B lifted LIBERO-Plus success from 50.06% to 56.82%.
Ropedia, founded in Singapore in the second half of 2025 by Zhaoxi Chen, Fangzhou Hong and Ziwei Liu, is an NTU spin-off with an office in Mountain View and an academic network spanning Princeton, CMU, UIUC, HK PolyU, Westlake and the Allen Institute for AI. The company positions itself as data infrastructure for physical AI; W0 is its bid to make one predictive model serve every embodiment.