*This dataset supports online use.Click here to jump.

OmniRetarget is a high-quality trajectory dataset for full-body motion remapping of humanoid robots, released in 2025 by Amazon, MIT, University of California, Berkeley and other institutions. The related paper results are "OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction".

This dataset contains the motion trajectories of the G1 humanoid robot as it interacts with objects and complex terrain, covering three scenarios: carrying objects, walking on terrain, and mixed object-terrain interaction. Due to licensing restrictions, the public dataset does not include the remapped version of LAFAN1. The dataset is divided into three subsets, totaling approximately four hours of motion trajectory data. The specific composition is as follows:

• robot-object: The trajectory of the robot carrying the object, derived from OMOMO 3.0 data;
• robot-terrain: The robot's motion trajectory on complex terrain, generated by internal MoCap, lasting about 0.5 hours;
• robot-object-terrain: A motion track involving both the object and the terrain, about 0.5 hours long.
• In addition, the dataset also contains a models directory, which provides visual model files in URDF, SDF, and OBJ formats for display rather than training.

The motion trajectory data is stored in .npz format. Each trajectory contains two fields: frame rate (fps) and system status (qpos).