This repository contains the source code of the paper LEO: Learning Energy-based Models in Factor Graph optimization.
The code contains three sets of experiments:
1. 1D energy-based regression: Learn E(x,y) where x is 1D input and y is 1D output
2. 2D synthetic navigation: Learn E(x,z) where x is a sequence of 2D robot poses and z is a sequence of GPS and odometry measurements.
3. 2D real-world pushing: Learn E(x,z) where x is a sequence of 2D object poses and z is a sequence of tactile measurements.
Create a virtual python environment using Anaconda:
conda create -n leo python=3.7
conda activate leo
Install the leopy python package locally. In the leopy/ dir execute:
pip install -e .
Any additional install requirements are detailed within each example.
We borrow this example from The Differentiable Cross-Entropy Method repository. We extend the example by adding in energy-based learning method (LEO), unrolling Gauss-Newton method and additional preprocessing and visualization utilities.
To run the regression example, execute in the leopy/ dir:
python scripts/examples/regression.py
By default, this runs LEO with Gauss-Newton as the inner loop optimizer. To run the example with other algorithms and settings, please take a look at regression.yaml. For instance, to run unrolled Gauss-Newton, set model=${unroll_gn}. To run LEO with cross-entropy method as the inner loop optimizer, set model=${leo_cem}.
Upon running the script, you should see the energy function E(x,y) being visualized as it is updated each iteration:
(green is the groundtruth function, orange are the LEO samples, lighter colors correspond to lower energy)
To run the navigation examples, we first need to install gtsam. Start by cloning the gtsam repository:
git clone https://github.com/borglab/gtsam.git
git checkout tags/4.0.0
Build and install the gtsam library:
cmake -DGTSAM_INSTALL_CYTHON_TOOLBOX=ON -DGTSAM_PYTHON_VERSION=3.7 ..
make -j
make install
If doing a local install, additionally pass in the install path -DCMAKE_INSTALL_PREFIX=../install.
Download the 2D navigation datasets locally by running:
./download_local_files.sh
To run the 2D navigation example:
python scripts/examples/nav2d.py
By default, this runs LEO with iSAM2 optimizer in the inner loop on the varying covariance dataset. To run the example with other datasets and settings, please take a look at leo_nav2d.yaml.
Upon running the script, you should see the optimized robot trajectories being saved to local/plots every learning iteration:
If you find this repository helpful in your publications, please consider citing the following:
@inproceedings{sodhi2021leo,
title={{LEO: Learning energy-based models in factor graph optimization}},
author={Sodhi, Paloma and Dexheimer, Eric and Mukadam, Mustafa and Anderson, Stuart and Kaess, Michael},
booktitle={Conference on Robot Learning (CoRL)},
year={2021}
}
This repository is licensed under the MIT License.