DemoGen

A method that automatically turns one real-robot demonstration into many synthetic demonstrations with objects in new positions.

DemoGen is a synthetic-demonstration-generation method from Huazhe Xu's group at Tsinghua University, in collaboration with the Shanghai Qi Zhi Institute and Shanghai AI Lab, published at RSS 2025. Imitation-learning policies generalize poorly across space — moving an object to a new position can make them fail — but collecting demonstrations by hand at every possible position is expensive. DemoGen needs only one human demonstration per task: it splits the trajectory into free-space transport segments and contact-based manipulation segments; when the object moves to a new position, the contact segment is transformed along with it, and the transport segment is reconnected using motion planning. Observations are produced by directly editing the point cloud, moving the points belonging to the object and the end effector together — no simulator is needed, and there's no need to re-collect data on the real robot. The paper reports generating a full dataset in about 22 seconds, versus roughly 83.7 hours for the MimicGen approach. Training pairs this with 3D Diffusion Policy (DP3), which takes point clouds as input.

ExampleOn a jar-opening task, with only 1 collected demonstration, DemoGen generates demonstrations with the jar in different positions; the paper reports the resulting policy reaches 100% success even at out-of-distribution positions.

Also called

DemoGen: Synthetic Demonstration Generation for Data-Efficient Visuomotor Policy Learning

Related

MimicGen · DexMimicGen · 3D Diffusion Policy · Synthetic Data · Spatial Generalization · Point Cloud

Sources

DemoGen 项目主页 (Chinese)
DemoGen: Synthetic Demonstration Generation for Data-Efficient Visuomotor Policy Learning (arXiv)

As of

2025-06