thaipduong / SE3HamDL

Code for our RSS'21 paper: "Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control"

Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control

This repo provides code for our paper "Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control". Please check out our project website for more details: https://thaipduong.github.io/SE3HamDL/.

Dependencies

Our code is tested with Ubuntu 18.04 and Python 3.7, Python 3.8. It depends on the following Python packages:

torchdiffeq 0.1.1, torchdiffeq 0.2.3

gym 0.18.0, gym 1.21.0

gym-pybullet-drones: https://github.com/utiasDSL/gym-pybullet-drones

torch 1.4.0, torch 1.9.0, torch 1.11.0

numpy 1.20.1

scipy 1.5.3

matplotlib 3.3.4

pyglet 1.5.27 (pendulum rendering not working with pyglet >= 2.0.0)

Notes: The NaN error during training with torch 1.10.0 or newer has been fixed!!!!!!!!!. However, training might be slower since we switch to float64. To use the float32 version with torch 1.9.0, run git checkout float32_tensors.

Demo with pendulum

Run python ./examples/pendulum/train_pend_SO3.py to train the model with data collected from the pendulum environment. It might take some time to train. A pretrained model is stored in ./examples/pendulum/data/pendulum-so3ham_ode-rk4-5p.tar

Run python ./examples/pendulum/analyze_pend_SO3.py to plot the generalized mass inverse M^-1(q), the potential energy V(q), and the control coefficient g(q)

Run python ./examples/pendulum/rollout_pend_SO3.py to verify that our framework respect energy conservation and SE(3) constraints by construction, and plots phase portrait of a trajectory rolled out from our dynamics.

Run python ./examples/pendulum/control_pend_SO3.py to test the energy-based controller with the learned dynamics.

Demo with quadrotor

Run python ./examples/quadrotor/train_quadrotor_SE3.py to train the model with data collected from the pybullet drone environment. It might take some time to train. A pretrained model is stored in ./examples/quadrotor/data/quadrotor-se3ham-rk4-5p.tar

Data collection

Run python ./examples/quadrotor/analyze_quadrotor_SE3.py to plot the generalized mass inverse M^-1(q), the potential energy V(q), and the control coefficient g(q)

Run python ./examples/quadrotor/rollout_quadrotor_SE3.py to verify that our framework respect energy conservation and SE(3) constraints by construction, and plots phase portrait of a trajectory rolled out from our dynamics.

Run python ./examples/quadrotor/control_quadrotor_SE3.py to test the energy-based controller with the learned dynamics.

Citation

If you find our papers/code useful for your research, please cite our work as follows.

1. T. Duong, N. Atanasov. Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control. RSS, 2021

@inproceedings{duong21hamiltonian,
author = {Thai Duong AND Nikolay Atanasov},
title = {{Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control}},
booktitle = {Proceedings of Robotics: Science and Systems},
year = {2021},
address = {Virtual},
month = {July},
DOI = {10.15607/RSS.2021.XVII.086} 
}