YanjieZe / 3D-Diffusion-Policy

[RSS 2024] 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations
https://3d-diffusion-policy.github.io
MIT License
445 stars 45 forks source link

Questions about shape-meta #52

Closed Selen-Suyue closed 2 months ago

Selen-Suyue commented 3 months ago

I encountered a problem when browsing Adroit_hammer.yaml. Could you tell me shape_meta: &shape_meta

acceptable types: rgb, low_dim

obs:

image:

shape: *image_shape

type: rgb

point_cloud: shape: [512, 3] type: point_cloud agent_pos: shape: [24] type: low_dim action: shape: [26] Are all actions here axis-angle motions? And what does agent_pos represent? Also, are the representations of different tasks the same?

YanjieZe commented 2 months ago

Hi, thank you for your interest.

Actions are joint positions for hand and end effector positions for arm.

Agent pos includes the similar information, such as joint positions and end effector positions.

Most tasks are the same. It is recommended to use joint positions and eef positions as actions. Please check Diffusion Policy for more details: https://diffusion-policy.cs.columbia.edu/