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

Two new AI frameworks, OmniRetarget and PSyGenTAB, have been developed to address challenges in robotics and medical AI. OmniRetarget preserves human-object interactions for humanoid robot training, while PSyGenTAB generates synthetic clinical data with privacy preservation.

OmniRetarget is a data generation engine that preserves human-object and human-environment interactions for humanoid whole-body locomotion and manipulation. It uses an interaction mesh to model spatial and contact relationships between an agent, terrain, and objects, generating kinematically feasible trajectories by minimizing Laplacian deformation between human and robot meshes^[1]. The engine was evaluated by retargeting motions from OMOMO, LAFAN1, and in-house MoCap datasets, generating over 8-hour trajectories. OmniRetarget achieved better kinematic constraint satisfaction and contact preservation than widely used baselines. Meanwhile, PSyGenTAB is a privacy-preserving framework for generating synthetic clinical tabular data via constrained optimization. It addresses the limitation of medical AI development due to limited access to high-quality clinical data. PSyGenTAB formulates synthetic healthcare data generation as a constrained optimization problem, preserving inter-feature clinical relationships and minority-class diagnostic patterns essential for reliable health AI^[2].