Reproduce “Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianqi Liu, Hanfei Zhao, Qifeng Liang, Hui En Pang, Jiarui Xu, Sayang Mu, Penghao Qiu, YingSheng Zhang and 7 more
Ropedia W0 predicts future egocentric video from an initial image, text, camera poses, and image-aligned skeletons, using human and robot video. On 400 evaluation windows across eight embodiments, its A14B model reaches 21.98 dB PSNR, versus 19.02 for Cosmos3-Nano. Ablations show 21.65 without camera-ray conditioning and 18.17 without skeletons; the interface models visual motion, not contact or object dynamics.
Key insights
- Projected skeletons allow the model to use each embodiment’s native kinematics without requiring a shared joint-space action vector.
- Human-video pretraining transfers useful prediction ability to robot settings, and performance generally improves as the amount of human pretraining data increases.
- Skeleton conditioning contributes strongly to prediction quality; camera conditioning has a smaller average effect but matters more when the viewpoint moves.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.