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
514 stars 50 forks source link

关于state和action具体元素的含义 #83

Closed Pepper-FlavoredChewingGum closed 2 weeks ago

Pepper-FlavoredChewingGum commented 2 weeks ago

您好,我正在执行bash scripts/gen_demonstration_adroit.sh hammer指令来生成专家数据。但是我是初学者,不太了解收集数据中state和action具体代表什么意思。比如这24维的state是机器人哪几个关节的pose?旋转和平移是以四元数还是什么格式来存储的?26维的action代表具体什么含义?是否包含位置和速度约束? 真诚期待您的回复! img shape: (1000, 84, 84, 3), range: [0, 255] point_cloud shape: (1000, 512, 6), range: [-0.2482137798412634, 255.0] depth shape: (1000, 84, 84), range: [0.6981372237205505, 2.1000139713287354] state shape: (1000, 24), range: [-1.2793629169464111, 1.6018180847167969] action shape: (1000, 26), range: [-0.9999999403953552, 1.0]

YanjieZe commented 2 weeks ago

Hi, thank you for your interest. We use position control. States and actions are all end-effector positions&rotations + dex hand joints. Rotation can be represented in the format you like, such as Euler angles and quaternion.