Event

Ropedia Releases Cross-Embodiment World Model W0, Boosting Robot Task Progress to 56.7%

2 reports · 2 sources

Singapore-based embodied intelligence startup Ropedia, together with Nanyang Technological University, Duke University, University of California San Diego, and Tsinghua University, released the cross-embodiment mobile manipulation world model Ropedia W0, which can predict first-person future scenes after humans, robotic arms, and humanoid robots perform actions. In real Unitree G1 mobile manipulation tasks, robot task progress increased from 33.3% to 56.7%. The model takes as input the initial frame, text description, camera motion trajectory, and hand actions to generate future first-person videos. The team used a multimodal data pipeline to build an approximately 13,000-hour Xp-Data training dataset; another report states it is approximately 10,000 hours of first-person human video. The model includes a 5B dense version and an A14B MoE version; A14B is used for cross-embodiment prediction evaluation, and 5B is used for real-time interaction and robot action learning experiments. The model adopts a shared camera-plus-skeleton representation, where one set of weights can predict the first-person view of humans, robotic arms, and humanoid robots such as Unitree G1 and Fourier GR-1; integrating a new robot only requires its URDF. After distillation, it runs at 21.3 FPS, can support real-time VR control, and can generate minute-level continuous mobile manipulation. The paper, code, weights, and demo have all been made public.