search

Genesis-Embodied-AI / RoboGen

A generative and self-guided robotic agent that endlessly propose and master new skills.
Apache License 2.0
549 stars 50 forks source link

scene population #25

Closed caichuang0415 closed 3 months ago

caichuang0415 commented 4 months ago

Thanks for sharing your wonderful work! I am eager to know where is the "scene population" codes in your project? It refers to the "text-to-image-to-3D generation" in your paper. Could you please point out that part's codes for me? Thank you very much!

yufeiwang63 commented 3 months ago

Hi,

Thanks for your interest in our work. For most rigid and articulated object manipulation & locomotion tasks, we used retrieval-based scene population, i.e., the objects are retrieved from Objaverse. For soft body manipulation tasks, we used the text-to-image-to-3d generation pipeline. Since soft body dynamics is not supported in pybullet, in this reproducing codebase we did not try to integrate this text-to-image-to-3d pipeline with pybullet.

However, if you are interested, we used this repo: https://github.com/ashawkey/stable-dreamfusion for image-2-3d generation.

Also, I believe the text23d generation field is evolving rapidly so there should be many more new repos these days that can do even better & faster generation.