CD-Lagrange Network

This is code for the paper https://arxiv.org/abs/2102.11206

Structure

This project has four different modules

• environments which contains classes to simulate the following environments:
  • Bouncing Ball
  • Pendulum
  • Newton Cradle
• experiments, which contains the concrete setups used in the dissertation. This includes the parameter configuration and the code for creating the plots used in this section.
• models contains the implementation of integrator analyzed in this work (CD-Lagrange) as well as the models it is compared to (residual network, VIN).
• utils contains the TRAIN function, which universally trains a network model given a specific environment, and the PREDICT function, which can be used to predict the trajectory of a given model and environment.

Usage

Install the python requirements with

conda env create -f environment.yml
conda activate cdl-net

Simulate experiments

To rerun the experiments from the paper, one can execute the main.py file, which will sequentially execute the pendulum, bouncing ball and newton's cradle experiment using the idealised touch feedback data regime. It is possible to access the models trained in every experiment by using dot-access:

from experiments import pendulum, newton_cradle

# Get the environment object of the pendulum experiment.
pendulum.env
# Get the cdl model of the pendulum experiment.
pendulum.cdl_model

# Get the environment object of the pendulum experiment.
newton_cradle.env
# Get the cdl model of the pendulum experiment.
newton_cradle.cdl_model

Create new experiments

To create a new experiment, one can directly import the modules (i.e., a jupyter notebook) and set the desired parameters

from environments import NewtonCradle
from models import CDLNetwork, ResNet
from utils import TRAIN, PREDICT

# Initialise the environment
env = NewtonCradle(steps=700, dt=0.01, epochs=1000)
# Generate environment data
env.generate()

# Train CD-LAGRANGE
cdl_model = TRAIN(env, CDLNetwork, name='CDL')

# Train RESNET
resnet = TRAIN(env, ResNet, name='ResNet')

# Predict trajectory given the initial state q0
cdl_data = PREDICT(env, cdl_model)
resnet_data = PREDICT(env, cdl_data)

In order to get a unique filename for the parameters used one can facilitate the function env.get_filename(name, suffix='png'). This is useful for saving plots, for example.

Acknowledgement

Thanks to Steindor, Alex, and Marc for their excellent support.