OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction

OmniRetarget is a novel data generation engine designed to enhance the retargeting of human motions for humanoid robots, addressing the embodiment gap by preserving critical spatial and contact relationships during locomotion and manipulation tasks. Utilizing an interaction mesh and minimizing Laplacian deformation between human and robot meshes, it generates kinematically feasible trajectories, achieving superior kinematic constraint satisfaction and contact preservation compared to existing methods. This advancement allows for efficient data augmentation across various robot embodiments and environments, enabling proprioceptive reinforcement learning policies to effectively perform complex skills, such as parkour, with minimal reward terms and no learning curriculum.