Festivus

policy

3D Diffusion Policy (DP3)

Shanghai Qizhi / SJTU / Tsinghua / Shanghai AI Lab · PyTorch

or hover any field below to flag it

Overview

Name
3D Diffusion Policy (DP3)
Author
Shanghai Qizhi / SJTU / Tsinghua / Shanghai AI Lab
Framework
PyTorch
License
mit
Skill type
manipulation
Evidence level
verified
Task description
Diffusion policy conditioned on compact 3D point cloud features. Extracts sparse point clouds (512-1024 points) from a single depth camera, encodes via lightweight point encoder, then generates action trajectories with a diffusion model. Strong spatial awareness enables precise manipulation and viewpoint generalization.

Spaces

Action space
joint-position · 7-dim · 10Hz
Observation space
  • type: multimodal
  • · point_cloud (512-1024 pts from RealSense L515)
  • · joint_pos

Links

HuggingFace repo
null
Paper (arXiv)
https://arxiv.org/abs/2403.03954

Compatible robots

0+1 mentioned but not in catalog yet

No robots list 3D Diffusion Policy (DP3) as compatible yet. Know of one? Flag it above.

Compatible environments

1
tabletop-cleannot in seed

Datasets that reference this policy

0

No datasets reference 3D Diffusion Policy (DP3) yet.