Reproduce “GRAIL: Generating Humanoid Loco-Manipulation from World Models” 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
Tianyi Xie, Haotian Zhang, Jinhyung Park, Zi Wang, Bowen Wen, Jiefeng Li, Xueting Li, Qingwei Ben and 12 more
GRAIL generates robot-compatible human–object interaction trajectories by placing a 3D object and robot-proportioned character in a specified scene, then generating and reconstructing motion. It uses these trajectories to train trackers and visual policies. Tracking success reached 88.9% in simulation; real-robot policies achieved 84% and 80% pick-up success on seen and unseen objects, and 90% stair-climbing success. These demonstrations require setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.