Multi-Agent Sensing

This package provides centralized implementations of Randomized Sequential Partitions (RSP) (see references) that are suitable for numerical simulations as well as a number of other algorithms for multi-agent planning via submodular maximization such as sequential planning and auction algorithms.

Additionally, we implemented two application scenarios

• A simple coverage problem. Agent centers are distributed across the unit square. Actions are disks near the agent centers.
• A mutual information based target tracking problem. Agents plan on (short) receding horizons via Monte Carlo tree search.

Installation

Clone this repository to my_path/MultiAgentSensing. Then, in order for Julia to find the SubmodularMaximization module that this package provides, we need to add this to the LOAD_PATH variable. To do so, edit .julia/config/startup.jl and add the line:

push!(LOAD_PATH, "$(homedir())/projects/MultiAgentSensing")

Dependencies

In addition to a number of registered Julia packages, some scripts for plotting data from experimental trials rely on RosDataProcess. Due to the target tracking code, this package also depends on HistogramFilters.jl. Neither package is registered with the Julia ecosystem, but both can be installed easily Pkg.add:

Pkg.add("https://github.com/mcorah/RosDataProcess")
Pkg.add("https://github.com/mcorah/HistogramFilters.jl")

Other dependencies

• Julia: POMDPs.jl, MCTS.jl
• Python: tikzplotlib or matplotlib2tikz: One of these should be installed with whichever Python installation that PyCall.jl is configured to load

Example code

This package includes simple sensor coverage problems as well as receding horizon target tracking and visual coverage problems.

For receding-horizon problems, press "enter" to advance the simulation.

• example/coverage/coverage_test.jl: This simulates run a static coverage problem and saves and image of the result. A group agents distributed on a rectangle can each choose to cover area under one of several nearby disks. The goal is to maximize the covered area.
• example/target_tracking/multi_robot_example.jl: This example visualizes a group of robots "tracking" moving targets by maintaining (common/centralized) Baye's filters based on noisy observations of target distance. The background is colored by probabilility density of target locations, averaged across targets
• example/target_tracking/multi_robot_coverage_example.jl: This example visualizes a group of robots seeking to "cover" moving targets by keeping them inside the sensing range of any one robot

References

If you use this package in published work, pleace consider citing either of the following:

For the coverage scenario and the initial implementation of RSP:

@inproceedings{corah2018cdc,
  author = {Corah, Micah and Michael, Nathan},
  title={Distributed Submodular Maximization on Partition Matroids for Planning
         on Large Sensor Networks},
  booktitle={Proc. of the {IEEE} Conf. on Decision and Control},
  year = {2018},
  month = dec,
  address = {Miami, FL},
}

Please cite the thesis for the target tracking scenario and everything else:

@phdthesis{corah2020phd,
  author = {Corah, Micah},
  title = {Sensor Planning for Large Numbers of Robots},
  school = {Carnegie Mellon University},
  year = {2020}
}