OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction
Researchers have developed OmniRetarget, a novel data generation engine designed to improve the training of humanoid robots. This system addresses limitations in existing methods by preserving crucial interactions between robots, their environment, and manipulated objects, which are often lost in translation from human motion capture data. OmniRetarget generates physically plausible and interaction-rich trajectories, enabling reinforcement learning policies to execute complex tasks like parkour and loco-manipulation on robots such as the Unitree G1 with fewer reward terms and no learning curriculum. AI
IMPACT Enables more efficient training of humanoid robots for complex physical tasks by improving data generation quality.