This repository contains numerical experiments that accompany the following publication:
Mihir Parmar*, Mathew Halm*, and Michael Posa. "Fundamental Challenges in Deep Learning for Stiff Contact Dynamics," IROS 2021.
- Python 3.6.9 or higher
- 16GB RAM
- Linux (tested on Ubuntu 18.04 LTS)
GPU is not necessary, although it will provide significant speedup for for the training process.
- numpy 1.19.5
- matplotlib 3.3.2
- PyTorch 1.7.0
- scipy 1.5.4
- sklearn 0.23.2
- argparse 1.1
A precomputed set of trajectories can be downloaded from here. However, the data-scripts directory contains the scripts to generate additional data and visualize die roll trajectories.
To generate more data, install MuJoCo 200 at ~/.mujoco/mjpro200, copy your license key to ~/.mujoco/mjkey.txt and move the cube_toss.xml that defines the system model to /.mujoco/model/.
To simulate the die roll system in MuJoCo and generate trajectories, use the desired contact settings by changing the stiffness and dampping values under the solref tag part of the xml script and then run:
python3 generate.py
This will result in 11,000 trajectories being stored at contactlearning/data/<stiffness_value_used>/
To visalize any trajectory using MuJoCo's rendering, run:
python3 visualize.py <path to trajectory>
Under contactlearning, RNNPredictor.py defines the model architecture used in the experiments and supports MLP, LSTM, GRU, and BiLSTM architectures.
To train a model on the set of generated trajectories with specific training settings, run:
python3 train.py <training-settings>
The possible arguments under <training-settings> include:
--stiffness <value>, the stiffness value from {2500, 300, 100} corresponding to which data is to used (default = 2500)--train_tosses <value>, number of training trajectories upto 10000 (default = 500)--tw <value>, history-length (default 16)--normalize, normalize the input data--batch_size <value>, batch-size to use during training (default = 64)--recurrent_mode <value>, DNN architecture from {mlp, lstm, gru, bilstm} to use (default = lstm)--lr <value>, learning-rate value to use with Adam Optimizer (default = 1e-4)--hidden_size <value>, width of the hidden-layer of the RNN (default = 256)--weight_decay <value>, weight-decay to use for regularization (default = 0)
Checkpoints updated after every epoch as well as the final trained models are stored in contactlearning/models/.
tensorboard logs from training are stored in contactlearning/Logs. Training and validation loss curves can be visualized at localhost:6006 using:
tensorboard --logdir=contactlearning/Logs
eval_utils.py includes helper methods used in train.py to evaluate trained models.
Both training and evaluation results are stored in a .json file at contactlearning/Results