InterScene: A Python repository from IDC-Flash

Synthesizing Physically Plausible Human Motions in 3D Scenes

Liang Pan · Jingbo Wang · Buzhen Huang · Junyu Zhang · Haofan Wang · Xu Tang · Yangang Wang
Southeast University Shanghai AI Laboratory Xiaohongshu Inc.

3DV 2024

Our framework enables physically simulated characters to perform long-term interaction tasks in diverse and complex 3D scenes via composing reusable skills that include sitting (gray), getting up (blue), and avoiding obstacles (red).

News

[Nov. 9, 2023] Code for training and evaluating the sit policy released.
[Oct. 16, 2023] Paper accepted to 3DV 2024. We plan to release the code in mid-November 2023.

Dependencies

Environment

To create the environment, follow the following instructions:

We recommend to install all the requirements through Conda by

conda create -n rlgpu python=3.7
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Download IsaacGym Preview 4 from the official site and install it via pip.

Dataset

To prepare data for training/evaluating InterCon (sit & get-up policies), follow the following instructions:

Download SMPL-X v1.1 from the official site. Put them in the body_models/smplx folder.
Download SAMP motion dataset from the official site. Put them in the samp folder. Please download Motion Clips (.pkl), which contains the SMPL-X parameters.
The file structure should look like this:

|-- InterScene
|-- body_models
    |-- smplx
        |-- SMPLX_FEMALE.npz
        |-- SMPLX_FEMALE.pkl
        |-- SMPLX_MALE.npz
        |-- ...
|-- samp
    |-- chair_mo_stageII.pkl
    |-- chair_mo001_stageII.pkl
    |-- chair_mo002_stageII.pkl
    |-- ...

Run the following script to generate reference motion dataset:

python InterScene/data/dataset_samp_sit/generate_motion.py --samp_pkl_dir ./samp --smplx_dir ./body_models/smplx

Run the following script to generate 3D object dataset:

python InterScene/data/dataset_samp_sit/generate_obj.py

Getting Started

InterCon (sit & get-up policies)

## training sit policy
python InterScene/run.py --task HumanoidLocationSit --cfg_env InterScene/data/cfg/humanoid_location_sit.yaml --cfg_train InterScene/data/cfg/train/rlg/amp_task_location_sit.yaml --motion_file InterScene/data/dataset_samp_sit/dataset_samp_sit.yaml --num_envs 4096 --headless

## evaluating sit policy
python InterScene/run.py --task HumanoidLocationSit --cfg_env InterScene/data/cfg/humanoid_location_sit.yaml --cfg_train InterScene/data/cfg/train/rlg/amp_task_location_sit.yaml --motion_file InterScene/data/dataset_samp_sit/dataset_samp_sit.yaml --num_envs 4096 --headless --checkpoint InterScene/data/models/policy_sit.pth --test

Citation

@inproceedings{pan2023synthesizing,
    title={Synthesizing Physically Plausible Human Motions in 3D Scenes}, 
    author={Liang Pan and Jingbo Wang and Buzhen Huang and Junyu Zhang and Haofan Wang and Xu Tang and Yangang Wang},
    booktitle={International Conference on 3D Vision (3DV)},
    year={2024}
}

References

This repository is built on top of the following amazing repositories:

Main code framework is from: ASE
Some scripts are from: Pacer, HuMoR

Please follow the license of the above repositories for usage of that part of the codebase.

IDC-Flash/InterScene