The following is a non-exhaustive and expanding list of dynamical systems with Discrete Morphological Symmetries. Each example can be reproduced in a 3D interactive environment running:
cd [...]/RobotEquivariantNN
python robot_symmetry_visualization.py robot=<robot> gui=True debug=False
There is a large number of real-world robotic systems with DMSs not listed in this library. For now re restrict
ourselves to the systems in the curated library of URDF configurations offered by robot_descriptions.py. Expect pull request to robot_descriptions.py to enlarge this list.
| Solo-C2 | Atlas | Bolt | A1 |
|---|---|---|---|
 |
 |
 |
 |
| Cassie | Baxter | HyQ-C2 | --- |
 |
 |
 |
------ |
| Solo | HyQ | ---- | ---- |
|---|---|---|---|
 |
 |
----- | ----- |
On the module nn/EquivariantModules.py you can find the BasisLinear) and Convolutional (BasisConv1d) layers classes.
These are the equivariant versions of the standard Pytorch Linear and Conv1D layers.
The parametrization of the equivariant layers (i.e., class signature) is identical to the unconstrained Pytorch layer classes.
The only additional parameters are the input output representations rep_in and rep_out.
Each representation class instance holds the Symmetry Group and the matrix representations of each of the group actions.
Each example of the library of equivariant dynamical systems holds the representations on the Euclidean Space Ed and
on the system join-space Q_J, required to construct the representation for any NN input output spaces.
Expect a tutorial for the use of these classes. For now, visit nn/ContactCNN and nn/EMLP for the classes of the
NN architectures used in the experiments of the paper.
The script used to run the two experiments described in the paper is train_supervised.py which receives configuration files using hydra.
The default configurations can be found on the cgf directory.
All experiments were run in an HPC using slurm. The slurm jobs are not made public but the following are some examples of how to run the experiments on standard machines
The training_supervised script will create an experiment directory on the project root folder, where the results of each individual run will be stored along with a tensorboard log file
holding several metrics logged during training and validation. After training is done, each model variant is evaluated on the test dataset and a log file test_metrics.csv is created
containing all evaluation metrics of the test set. These test_metrics.csv files for each of the model variants and individual seeds are used by the scripts in the paper directory
to generate several figures including the ones presented in the paper.
- Atlas: Example experiment training
mlpandemlpmodels with256,512, and1024hidden channel neurons, with and without data augmentation.
python train_supervised.py --multirun robot=Atlas dataset=com_momentum dataset.samples=500000 dataset.augment=True,False model.num_channels=256,512,1024 model=emlp,mlp dataset.train_ratio=0. model.lr=1.5e-3 model.inv_dims_sca exp_name=com_sample_eff$
- Solo with Morphological Symmetry Group C2 and K4
python train_supervised.py --multirun robot=Solo,Solo-c2 dataset=com_momentum dataset.augment=True,False dataset.train_ratio=0.70 model=emlp,mlp model.lr=2.4e-3 model.num_channels=64,128,256,512 exp_name=com_sample_eff
- CNN
python train_supervised.py --multirun robot=mini-cheetah-c2 dataset=contact dataset.data_folder=training_splitted dataset.train_ratio=0.85 dataset.augment=True,False exp_name=contact_sample_eff_splitted model=contact_cnn model.lr=1e-4
- ECNN
python train_supervised.py --multirun robot=mini-cheetah-c2 dataset=contact dataset.data_folder=training_splitted dataset.train_ratio=0.85 dataset.augment=False exp_name=contact_sample_eff_splitted model=contact_ecnn model.lr=1e-5