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

Ropedia W0. Scaling Loco-Manipulation World Models across Embodiments with Human Experience. By Tianqi Liu and fourteen others, from Nanyang Technological University, Ropedia, and three other institutions. Published on the nineteenth of October. Its code repository is worked out: about sixty-five source files, with a readme and a licence.

A world model lets an embodied agent anticipate what its actions will do. Doing that for locomotion and manipulation together, and across many different bodies, is hard for two reasons the authors name. The action spaces differ from one embodiment to the next, and multimodal supervision is limited. A person and a robot, or two different robots, do not act through the same set of controls, so a model trained on one has no obvious way to learn from another. And the paired, annotated recordings that would let it try are scarce.

Ropedia W0 is presented as the first egocentric loco-manipulation world foundation model to span both humans and robots of diverse embodiments. The bridge between those bodies is a unified kinematic representation. Locomotion is encoded through camera poses, and manipulation through projected skeletons. For robots, the skeletons combine embodiment-specific forward kinematics with extensive manual keypoint annotation and correction.

The data behind it is Xp-Data, which pairs ten thousand hours of egocentric human data with three thousand hours of robot data, under unified multimodal annotations and systematic quality control. The authors also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark built from high-quality dexterous humanoid teleoperation data. The proportions are telling. Human footage outweighs robot footage by more than three to one, which is only useful if the representation lets the human hours inform the robot ones. That is the job the shared kinematic encoding is there to do.

The model reports state-of-the-art performance on five cross-embodiment benchmarks: DreamDojo, psi zero, omega home, Xp-Teleop, and AgiBot. It does the same on held-out human egocentric video, in settings ranging from zero-shot inference through few-shot and full-data post-training. Scaling up the human pre-training data improves results on both human and robot evaluations. After distillation, the model supports real-time, minute-long rollouts, enough to allow live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling are offered as further evidence of its use for embodied decision-making.

The scaling result carries the title's claim. If more human video helps a robot evaluation, then human experience is doing work for machines, and the shared kinematic language of camera poses and skeletons is what lets it cross over.