Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

Researchers unveil Ropedia W0, a world model for embodied agents that jointly models locomotion and manipulation across diverse human and robotic embodiments. According to the announcement, the team curated Xp-Data, a dataset totaling 10,000 hours of human data and 3,000 hours of robot data, featuring unified multimodal annotations and systematic quality control. The model employs a unified kinematic representation to encode locomotion via camera poses and manipulation through projected skeletons, combining forward kinematics for robots and extensive manual keypoint annotation and correction for skeletons.

Ropedia W0 achieves state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, Ψ_0, ω-HOME, Xp-Teleop, and AgiBot) as well as on held-out human egocentric videos, with notable improvements after pre-training on human data. After distillation, the model supports real-time, minute-long rollouts, enabling live VR interaction and applications in model-based planning and policy evaluation for embodied decision-making.