Skip to content

Latest commit

 

History

History
96 lines (68 loc) · 3.89 KB

README.md

File metadata and controls

96 lines (68 loc) · 3.89 KB
Raw

Subspace Graph Physics:
Accurate and Real-Time Physics Simulation Approach

[Paper]

Subspace Graph Physics is an accurate, real-time engineering approach for large-scale 3D/2D physics simulations. This is a computationally efficient version of "Learning To Simulate" developed by DeepMind. This approach can work with a single desktop GPU (with ~10GB vRAM) in training and can perform 2-3 orders of magnitude faster than physics-based methods (with similar accuracy) in inference on CPU.

drawing

Highlights

  • We train the graph network (GN) model in subspace by performing Principal Component Analysis (PCA).

  • PCA enables GN to be trained using a single desktop GPU with moderate VRAM for large 3D configurations.

  • The training datasets can be generated by our efficient and accurate Material Point Method (MPM).

  • The rollout runtime is under 1 sec/sec, and the training runtime is 60 global-step/sec (on NVIDIA RTX 3080).

  • The particle positions and velocities, and rigid body interaction forces are compared in the video above.

Install and Run Demo

  • Install

  • Train

    python -m learning_to_simulate.train --mode=train --eval_split=train --batch_size=2 --data_path=./learning_to_simulate/datasets/Excavation_PCA --model_path=./learning_to_simulate/models/Excavation_PCA

  • Test

    python -m learning_to_simulate.train --mode=eval_rollout --eval_split=test --data_path=./learning_to_simulate/datasets/Excavation_PCA --model_path=./learning_to_simulate/models/Excavation_PCA --output_path=./learning_to_simulate/rollouts/Excavation_PCA

  • Visualize

    • 2D plot

      python -m learning_to_simulate.render_rollout_2d_force --plane=xy --data_path=./learning_to_simulate/datasets/Excavation_PCA --rollout_path=./learning_to_simulate/rollouts/Excavation_PCA/rollout_test_0.pkl

    • 3D plot

      python -m learning_to_simulate.render_rollout_3d --fullspace=True --data_path=./learning_to_simulate/datasets/Excavation_PCA --rollout_path=./learning_to_simulate/rollouts/Excavation_PCA/rollout_test_0.pkl

Bibtex

Please cite our papers [1, 2] if you use this code for your research:

@article{HAERI2024108765,
   title = {Subspace graph networks for real-time granular flow simulation with applications to machine-terrain interactions},
   journal = {Engineering Applications of Artificial Intelligence},
   volume = {135},
   pages = {108765},
   year = {2024},
   issn = {0952-1976},
   doi = {https://doi.org/10.1016/j.engappai.2024.108765},
   url = {https://www.sciencedirect.com/science/article/pii/S0952197624009230},
   author = {Amin Haeri and Daniel Holz and Krzysztof Skonieczny},
   keywords = {Real-time physics simulation, Geometric deep learning, Graph neural networks, Continuum mechanics, Experiment},
}

and/or

@INPROCEEDINGS{9438132,
    author={Haeri, A. and Skonieczny, K.},
    booktitle={2021 IEEE Aerospace Conference (50100)},
    title={Accurate and Real-time Simulation of Rover Wheel Traction},
    year={2021},
    volume={},
    number={},
    pages={1-9},
    doi={10.1109/AERO50100.2021.9438132}
}