iDP3

An improved 3D Diffusion Policy that lets a humanoid trained on data from a single scene generalize to new scenes.

iDP3 was released in October 2024 by Yanjie Ze, Jiajun Wu, and colleagues at Stanford, together with Simon Fraser University, UPenn, UIUC, and CMU, published at IROS 2025, an improved version of 3D Diffusion Policy (DP3). The original DP3 processes point clouds in world coordinates, which requires camera calibration and cropping the point cloud by segmentation — not a good fit for a humanoid whose head camera moves along with its body. iDP3 switches to an “egocentric” 3D point cloud in camera coordinates instead, eliminating the need for calibration and segmentation, while also scaling up the input point-cloud size, improving the vision encoder, and lengthening the action-prediction horizon. The accompanying system includes upper-body teleoperation based on an Apple Vision Pro, and a 25-degree-of-freedom Fourier GR1 humanoid platform (with a RealSense L515 depth camera) mounted on a height-adjustable cart. Using data collected from just a single scene, the robot can complete tasks in many new scenes using only its onboard compute.

ExampleA policy trained only on demonstration data collected in one scene can, once deployed, perform the same kind of manipulation in other real-world scenes it has never seen.

Also called

Improved 3D Diffusion Policy, Generalizable Humanoid Manipulation with 3D Diffusion Policies

Related

3D Diffusion Policy · Diffusion Policy · Point Cloud · Humanoid Robot · Fourier GR-1 · VR Teleoperation

Sources

Generalizable Humanoid Manipulation with 3D Diffusion Policies (arXiv 2410.10803)
Project page
YanjieZe/Improved-3D-Diffusion-Policy (GitHub)

As of

2025-09