SGP-based Optimization Tools

Software Suite for Sensor Placement (SP) and Informative Path Planning (IPP).

The library includes python code for the following:

• Greedy algorithm-based approaches
• Bayesian optimization-based approaches
• Genetic algorithm-based approaches
• Sparse Gaussian process (SGP)-based approaches

       

Installation

The library is available as a pip package. To install the package, run the following command:

python3 -m pip install sgptools

Installation from source:

git clone https://github.com/itskalvik/sgp-tools.git
cd sgp-tools/
python3 -m pip install -r requirements.txt
python3 -m pip install -e .

Note: The requirements.txt file contains packages and their latest versions that were verified to be working without any issues.

Quick Start

Please refer to the examples folder for Jupyter notebooks demonstrating all the methods included in the library 😄

Method Summary

Codemap

• examples/: Jupyter notebooks with code to demonstrate each method in the library
  • IPP.ipynb: SGP-based IPP (point, non-point, continuous sensing, distance constrained, and multi-robot)
  • IPPBaselines.ipynb: SGP-based IPP approach and baseline methods
  • non_point_FoV.ipynb: IPP with non-point FoV sensors (drone camera setup)
  • non_stationary_demo.ipynb: SP with non-stationary kernel
  • obstacles.ipynb: SP in an environment with obstacles
• sgptools/: SGP-Tools library
  • kernels/: Kernel functions
    • neural_kernel.py: Neural Non-Stationary Spectral Kernel
  • models/: Sensor placement and IPP methods
    • core/: GP/SGP models used for sensor placement and IPP
      • augmented_gpr.py: GPflow's GP that supports transforms (expansion and aggregation)
      • augmented_sgpr.py: GPflow's SGP that supports transforms (expansion and aggregation)
      • transformations.py: Expansion and aggregation transforms for IPP
      • osgpr.py: Thang Bui's implementation of online sparse variational GP used for online/adaptive IPP
    • bo.py: Bayesian optimization-based sensor placement method that maximizes mutual information
    • cma_es.py: Genetic algorithm-based sensor placement method that maximizes mutual information
    • continuous_sgp.py: Continuous SGP-based sensor placement method
    • greedy_mi.py: Greedy sensor placement method that maximizes mutual information
    • greedy_sgp.py: Greedy SGP-based sensor placement method
  • utils/: Tools used for preprocessing the data, training GPs and SGPs, and generating paths
    • data.py: Tools to preprocess datasets
    • gpflow.py: Tools to interface with GPflow
    • metrics.py: Metrics used to quantify the solution quality
    • misc.py: Miscellaneous helper functions
    • tsp.py: TSP solver

Datasets

• High-resolution topography data can be downloaded from NOAA Digital Coast

• High-resolution bathymetry data can be downloaded from NOAA Digital Coast

• Large tif files need to be downsampled using the following command (requires GDAL package):

  gdalwarp -tr 50 50 <input>.tif <output>.tif

About SGP-Tools

Please consider citing the following papers if you use SGP-Tools in your academic work 😄

@misc{JakkalaA23SP,
AUTHOR={Kalvik Jakkala and Srinivas Akella},
TITLE={Efficient Sensor Placement from Regression with Sparse Gaussian Processes in Continuous and Discrete Spaces},
NOTE= {Preprint},
YEAR={2023},
URL={https://itskalvik.github.io/publication/sgp-sp},
}

@inproceedings{JakkalaA24IPP,
AUTHOR={Kalvik Jakkala and Srinivas Akella},
TITLE={Multi-Robot Informative Path Planning from Regression with Sparse Gaussian Processes},
booktitle={IEEE International Conference on Robotics and Automation, {ICRA}},
YEAR={2024},
PUBLISHER = {{IEEE}},
URL={https://itskalvik.github.io/publication/sgp-ipp}
}

Acknowledgements

This work was funded in part by the UNC Charlotte Office of Research and Economic Development and by NSF under Award Number IIP-1919233.

License

The SGP-Tools software suite is licensed under the terms of the Apache License 2.0. See LICENSE for more information.