Sample efficient inverse dynamics learning

Inverse dynamics Learning of Magneto

The code is developed based on

• graph_nets from deepmind
• dartpy for dynamics simulation environment

Citation

@article{lee2022sample,
  title={Sample Efficient Dynamics Learning for Symmetrical Legged Robots: Leveraging Physics Invariance and Geometric Symmetries},
  author={Lee, Jee-eun and Lee, Jaemin and Bandyopadhyay, Tirthankar and Sentis, Luis},
  journal={arXiv preprint arXiv:2210.07329},
  year={2022}
}

1. Virtual environment settings

Due to the current issue in using darpy with conda, please use python3-venv.

Python3-venv setup

Install python3-venv package that provides the venv module.

$ sudo apt install python3-venv

Within the directory run the following command to create your new virtual environment: (just replace "path/to/workspace" to your workspace path)

$ cd path/to/workspace
$ mkdir venv
$ python3 -m venv venv1

Activate the created venv

$ source venv1/bin/activate

1.[Optional-just for simulator] Install dartpy

Please make sure to install all the dependancies before installing dartpy.

Details can be found here

2-1 Pip list for auto install

(CPU)

$ pip install -r path/to/workspace/requirements.txt

2-2 Pip list for manual install

• (CPU)

$ pip install graph_nets "tensorflow>=2.1.0-rc1" "dm-sonnet>=2.0.0b0" tensorflow_probability

• (GPU)

$ pip install graph_nets "tensorflow_gpu>=2.1.0-rc1" "dm-sonnet>=2.0.0b0" tensorflow_probability

$ pip install PyGeometry
$ pip install scipy
$ pip install urdfpy

[optional]

$ pip install dartpy

2.Code Explanation

• a_dataset: raw dataset / tf dataset are saved here
• a_result: learned model, validation errors, logs are saved here
• ffnn_inverse_dynamics: all codes for feed forward neural network
• typical_gnn_inverse_dynamics: all codes for typical graph neural network
• gn_inverse_dynamics: all codes for symetry considered graph neural network
• sn_inverse_dynamics: all codes for symetry considered nerual network

1. tfData generation

Dataset generation for tensorflow use, in each model representation (e.g. ffnn-vector tensor, tgnn,gnn-graph tuple tensor, sn-group of vectors tensor)

• script : run a script <a_scripts/run_dataset>

  source a_scripts/run_dataset
  

• runfile:

  ffnn: ffnn_inverse_dynamics/hexamagneto/main_ffnn_make_dataset.py

  tgnn: typical_gnn_inverse_dynamics/robot_graph_generator/hexamagneto/main_make_dataset_gn.py

  sn(proposed):sn_inverse_dynamics/hexamagneto/main_sn_make_dataset.py

  gnn(proposed):gn_inverse_dynamics/robot_graph_generator/hexamagneto/main_make_dataset_gn.py

2. train the model / calculate test error for other data

• script:

  magneto: a_scripts/run_test_(ffnn/gn/tgn/sn)

  hexa-magneto: a_scripts/run_test_(ffnn/gn/tgn/sn)_hexa

• runfile for train

  main_train_model_(ffnn/gn/tgn/sn).py

• runfile for generalization error

  main_check_model_(ffnn/gn/tgn/sn).py