Siyuan Huang*, Zan Wang*, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, Song-Chun Zhu
This repository is the official implementation of paper "Diffusion-based Generation, Optimization, and Planning in 3D Scenes".
We introduce SceneDiffuser, a conditional generative model for 3D scene understanding. SceneDiffuser provides a unified model for solving scene-conditioned generation, optimization, and planning. In contrast to prior work, SceneDiffuser is intrinsically scene-aware, physics-based, and goal-oriented.
Paper | arXiv | Project | HuggingFace Demo | Checkpoints
We introduce SceneDiffuser, a conditional generative model for 3D scene understanding. SceneDiffuser provides a unified model for solving scene-conditioned generation, optimization, and planning. In contrast to prior works, SceneDiffuser is intrinsically scene-aware, physics-based, and goal-oriented. With an iterative sampling strategy, SceneDiffuser jointly formulates the scene-aware generation, physics-based optimization, and goal-oriented planning via a diffusion-based denoising process in a fully differentiable fashion. Such a design alleviates the discrepancies among different modules and the posterior collapse of previous scene-conditioned generative models. We evaluate SceneDiffuser with various 3D scene understanding tasks, including human pose and motion generation, dexterous grasp generation, path planning for 3D navigation, and motion planning for robot arms. The results show significant improvements compared with previous models, demonstrating the tremendous potential of SceneDiffuser for the broad community of 3D scene understanding.
- [ 2023.04 ] We release the code for grasp generation and arm motion planning!
-
Create a new
condaenvironemnt and activate it.conda create -n 3d python=3.8 conda activate 3d
-
Install dependent libraries with
pip.pip install -r pre-requirements.txt pip install -r requirements.txt
- We use
pytorch1.11andcuda11.3, modifypre-requirements.txtto install other versions ofpytorch.
- We use
-
Install Isaac Gym and install pointnet2 by executing the following command (optional for grasp generation and arm motion planning).
pip install git+https://github.com/daveredrum/Pointnet2.ScanNet.git#subdirectory=pointnet2
You can use our pre-processed data or process the data by yourself following the instructions.
But, you also need to download some official released data assets which are not processed, see instructions. Please remember to use your own data path by modifying the path configuration in:
-
scene_model.pretrained_weightsinmodel/*.yamlfor the path of pre-trained scene encoder (if you use a pre-trained scene encoder) -
dataset.*_dir/dataset.*_pathconfigurations intask/*.yamlfor the path of data assets
Download our pre-trained model and unzip them into a folder, e.g., ./outputs/.
| task | checkpoints | desc |
|---|---|---|
| Pretrained Point Transformer | 2022-04-13_18-29-56_POINTTRANS_C_32768 | |
| Pose Generation | 2022-11-09_11-22-52_PoseGen_ddm4_lr1e-4_ep100 | |
| Motion Generation | 2022-11-09_12-54-50_MotionGen_ddm_T200_lr1e-4_ep300 | w/o start position |
| Motion Generation | 2022-11-09_14-28-12_MotionGen_ddm_T200_lr1e-4_ep300_obser | w/ start position |
| Path Planning | 2022-11-25_20-57-28_Path_ddm4_LR1e-4_E100_REL | |
| Grasp Generation | 2022-11-15_18-07-50_GPUR_l1_pn2_T100 | |
| Arm Motion Planning | 2022-11-11_14-28-30_FK2Plan_ptr_T30_4 | denoising step is 30 |
-
Train with single gpu
bash scripts/pose_gen/train.sh ${EXP_NAME} -
Train with 4 GPUs (modify
scripts/pose_gen/train_ddm.shto specify the visible GPUs)bash scripts/pose_gen/train_ddm.sh ${EXP_NAME}
bash scripts/pose_gen/test.sh ${CKPT} [OPT]
# e.g., bash scripts/pose_gen/test.sh ./outputs/2022-11-09_11-22-52_PoseGen_ddm4_lr1e-4_ep100/ OPT[OPT]is optional for optimization-guided sampling.
bash scripts/pose_gen/sample.sh ${CKPT} [OPT]
# e.g., bash scripts/pose_gen/sample.sh ./outputs/2022-11-09_11-22-52_PoseGen_ddm4_lr1e-4_ep100/ OPT[OPT]is optional for optimization-guided sampling.
The default configuration is motion generation without observation. If you want to explore the setting of motion generation with start observation, please change the task.has_observation to true in all the scripts in folder ./scripts/motion_gen/.
-
Train with single gpu
bash scripts/motion_gen/train.sh ${EXP_NAME} -
Train with 4 GPUs (modify
scripts/motion_gen/train_ddm.shto specify the visible GPUs)bash scripts/motion_gen/train_ddm.sh ${EXP_NAME}
bash scripts/motion_gen/test.sh ${CKPT} [OPT]
# e.g., bash scripts/motion_gen/test.sh ./outputs/2022-11-09_12-54-50_MotionGen_ddm_T200_lr1e-4_ep300/ OPT[OPT]is optional for optimization-guided sampling.
bash scripts/motion_gen/sample.sh ${CKPT} [OPT]
# e.g., bash scripts/motion_gen/sample.sh ./outputs/2022-11-09_12-54-50_MotionGen_ddm_T200_lr1e-4_ep300/ OPT[OPT]is optional for optimization-guided sampling.
To run this code, you first need to change the git branch to obj by executing
git checkout objMake sure you have installed Isaac Gym and pointnet2. See Setup section.
-
Train with single gpu (one gpu is enough)
bash scripts/grasp_gen_ur/train.sh ${EXP_NAME}
bash scripts/grasp_gen_ur/sample.sh ${CKPT} [OPT]
# e.g., bash scripts/grasp_gen_ur/sample.sh ./outputs/2022-11-15_18-07-50_GPUR_l1_pn2_T100/ OPT[OPT]is optional for optimization-guided sampling.
You first need to run scripts/grasp_gen_ur/sample.sh to sample some results. Then we will compute quantitative metrics with these sampled results.
bash scripts/grasp_gen_ur/test.sh ${EVAL_DIR} ${DATASET_DIR}
# e.g., bash scripts/grasp_gen_ur/test.sh outputs/2022-11-15_18-07-50_GPUR_l1_pn2_T100/eval/final/2023-04-20_13-06-44 YOUR_PATH/data/MultiDex_UR-
Train with single gpu
bash scripts/path_planning/train.sh ${EXP_NAME} -
Train with 4 GPUs (modify
scripts/path_planning/train_ddm.shto specify the visible GPUs)bash scripts/path_planning/train_ddm.sh ${EXP_NAME}
bash scripts/path_planning/plan.sh ${CKPT}
# e.g., bash scripts/path_planning/plan.sh ./outputs/2022-11-25_20-57-28_Path_ddm4_LR1e-4_E100_REL/bash scripts/path_planning/sample.sh ${CKPT} [OPT] [PLA]
# e.g., bash scripts/path_planning/sample.sh ./outputs/2022-11-25_20-57-28_Path_ddm4_LR1e-4_E100_REL/ OPT PLA- The program will generate trajectories with given start position and scene; rendering the results into images. (The results not the planning results, just use diffuser to generate diverse trajectories.)
[OPT]is optional for optimization-guided sampling.[PLA]is optional for planner-guided sampling.
To run this code, you first need to change the git branch to obj by executing
git checkout objMake sure you have installed Isaac Gym and pointnet2. See Setup section.
-
Train with single gpu
bash scripts/franka_planning/train.sh ${EXP_NAME} -
Train with 4 GPUs (modify
scripts/path_planning/train_ddm.shto specify the visible GPUs)bash scripts/franka_planning/train_ddm.sh ${EXP_NAME}
bash scripts/franka_planning/plan.sh ${CKPT}
# e.g., bash scripts/franka_planning/plan.sh outputs/2022-11-11_14-28-30_FK2Plan_ptr_T30_4/If you find our project useful, please consider citing us:
@article{huang2023diffusion,
title={Diffusion-based Generation, Optimization, and Planning in 3D Scenes},
author={Huang, Siyuan and Wang, Zan and Li, Puhao and Jia, Baoxiong and Liu, Tengyu and Zhu, Yixin and Liang, Wei and Zhu, Song-Chun},
journal={arXiv preprint arXiv:2301.06015},
year={2023}
}Some codes are borrowed from latent-diffusion, PSI-release, Pointnet2.ScanNet, point-transformer, and diffuser.
This project is licensed under the MIT License. See LICENSE for more details. The following datasets are used in this project and are subject to their respective licenses:
- PROX is under the Software Copyright License for non-commercial scientific research purposes.
- LEMO is under the MIT License.
- ScanNet V2 is under the ScanNet Terms of Use.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent