Multi Agent NMPC for Active Information Acquisition

h1. Multi Agent NMPC for Active Information Acquisition wiki

h2. Foreword

This wiki is not entirely up-to-date, but will be very soon.

In case of a question, query, comment, or bug report, feel free to use the issues system or contact the corresponding author Martin Jacquet - martin.jacquet@laas.fr.

>

h2. Prerequisite

The framework has been written and tested using *Ubuntu 18.04* since it is the OS used by the LAAS-CNRS robotic platform. It should work seamlessly on a recent Linux version, but there is no guarantee.

Some issues have been found while installing the software on Ubuntu 16.04 because of version incompatibility with Protoc and Protobuf.

The installation on a non-Linux OS has to be handled by the user.

>

The installation assumes the use of a package manager (e.g. @apt@) to install some dependencies, as well as the Gazebo simulator. Everything provided in this repo or by the LAAS-CNRS robotic platform aims to be installed locally in the repository folder to avoid polluting the user's system.

>

h2. I - Software Overview

h3. I.1. Openrobots

Collections of all the open-source software used at LAAS. You can find more details in "Openrobots Wiki-Homepage":https://www.openrobots.org/wiki

>

h3. I-2. Robotpkg

"Robotpkg":http://robotpkg.openrobots.org/ is a packaging system for installing robotics software developed by the robotic community.

We will use robotpkg to install the required modules for the simulations (state estimation, gazebo interface...) as well as third-party dependencies (qpOases).

>

h3. I-3. GenoM

GenoM is a generator of modules, designed to be middleware independent, i.e. the same module can be compiled for, e.g., ROS, YARP, or Pocolibs, without any modification.

This allows a great code re-usability and abstracts the user from any specific choice of middleware.

Originally GenoM has been developed tightly with Pocolibs, then from version 3, aka GenoM3, ROS templates have been provided.

>

Another specificity of GenoM is the interaction with and between components.

Each component is started independently like a Linux executable (within a roscore, for ROS, or an h2 instance, for Pocolibs), then the connection between ports (or topics) is made using a supervisor, "Genomix":https://git.openrobots.org/projects/genomix, either with "Matlab":https://git.openrobots.org/projects/matlab-genomix or "TCL":https://git.openrobots.org/projects/tcl-genomix. 

>

h3. I-4. Pocolibs

"Pocolibs":https://www.openrobots.org/wiki/pocolibs/ is a middleware, like ROS.

It aims at being lighter and faster than ROS, when running on a single machine, thanks to the exploitation of shared memory. ROS, on the other hand, uses a network layer for sending messages between nodes, this leads to greater delays and loss of performance.

>

h3. I-5. TeleKyb

The "TeleKyb":https://git.openrobots.org/projects/telekyb3 software platform provides the aerial-robotic oriented software developed at LAAS-CNRS.

In particular, we will use:

* "mrsim":https://git.openrobots.org/projects/mrsim-genom3, a Multi-Robot SIMulator. It is designed to be a transparent interface w.r.t. the real aerial vehicles used in LAAS-CNRS. It makes the transition between simulation and experiments transparent, from the software point of view.

* "pom":https://git.openrobots.org/projects/pom-genom3, a UKF-based state estimator merging state feedback for different sources (e.g. mocap + IMU)

* "optitrack":https://git.openrobots.org/projects/optitrack-genom3, to export the motion capture data to the genom software stack

* "rotorcraft":https://git.openrobots.org/projects/rotorcraft-genom3, the low-level interface, with either the simulated or real platform

* "nhfc":https://git.openrobots.org/projects/nhfc-genom3, near-hovering flight controller, used for unmodeled take-off and post-failure recoveries

* "maneuver":https://git.openrobots.org/projects/maneuver-genom3, a global trajectory planner, providing position and attitude (as quaternions) as well as first and second derivatives. It implements take-off and waypoint-to-waypoint motions.

>

h3. I-6. Gazebo

To simulate the platform, we use the "Gazebo":http://gazebosim.org/ simulator. To interface it with the genom software stack, we use two dedicated components:

* "mrsim-gazebo":https://git.openrobots.org/projects/mrsim-gazebo a plugin to interface the simulated multi-rotor with the genom components (in place of mrsim)

* "optitrack-gazebo":https://git.openrobots.org/projects/optitrack-gazebo emulates the optitrack network interface to publish the model poses

>

The installation procedure for Gazebo can be found at http://www.gazebosim.org/tutorials?cat=install&tut=install_ubuntu&ver=9.0

h2. II - Installation procedure

This section is a tutorial on how to install the software architecture to run the simulations.

>

h3. II-0. Clone the Perceptive and torque-control NMPC repository

Clone the repo associated to this project. Its root will act as the devel folder for the following.

<pre><code class="shell">

git clone git://redmine.laas.fr/laas/perceptive-torque-nmpc.git

cd ./perceptive-torque-nmpc/

</code></pre>

>

To simplify the installation, we provide some environment variables in the @env.sh@ file.

In order to run all the installed executables, we need to set up the path to the newly created folders.

We provide a @env.sh@ script that exports all the required variables.

*/!\* the source has to be called in the repository root since it uses the @pwd@ command to export the paths.

<pre><code class="shell">

source env.sh

</code></pre>

>

h3. II-1. Setup robotpkg

(Steps taken from http://robotpkg.openrobots.org/install.html)

h4. 1. Clone the robotpkg lastest release:

<pre><code class="shell">

git clone git://git.openrobots.org/robots/robotpkg

</code></pre>

h4. 2. Check that the @openrobots/@ folder exists in the repository root, and update the environement variables accordingly if you didn't source the @env.sh@ file:

<pre><code class="shell">

export ROBOTPKG_BASE=`pwd`/openrobots

</code></pre>

h4. 3. Install robotpkg

<pre><code class="shell">

cd robotpkg/bootstrap

./bootstrap --prefix=$ROBOTPKG_BASE

</code></pre>

h4. 4. Install the required components and their dependencies

The installation can be done 'manually' by navigating to the desired folder in @./robotpkg/@ and installing with @make update@; but we will simplify the process using a _set_.

To do so, we need to edit the config file: @$ROBOTPKG_BASE/etc/robotpkg.conf@. Add the following at the end of the file:

<pre><code class="shell">

PKG_OPTIONS.%-genom3 = \

        codels \

        pocolibs-server \

        pocolibs-client-c

PKGSET.mpcset = \

    sysutils/arduio-genom3 \

    architecture/genom3 \

    architecture/genom3-pocolibs \

    robots/rotorcraft-genom3 \

    localization/pom-genom3 \

    localization/optitrack-genom3 \

    motion/nhfc-genom3 \

    path/libkdtp \

    optimization/qpoases \

    net/genomix \

    supervision/matlab-genomix \

    supervision/tcl-genomix \

    shell/eltclsh \

    simulation/mrsim-genom3 \

    simulation/mrsim-gazebo \

    simulation/libmrsim \

    simulation/optitrack-gazebo

PREFER.lapack = robotpkg

PREFIX.matlab = <path/to/Matlab>

</code></pre>

The last line needs to point to the Matlab root folder in the system (e.g. @/opt/Matlab@).

It is recommended to use Matlab for the proposed simulations since the syntax is more intuitive and comprehensible for the user to modify. However, we also provide all the launch files in TCL, as well as the environment to run them (@shell/eltclsh@ in the above list is a custom TCL script shell).

If Matlab is not installed on the system, remove the lines @supervision/matlab-genomix \@ and @PREFIX.matlab = <path/to/Matlab>@ from the above list.

Also, all the above is meant for using Pocolibs, not ROS. Futur version of this tutorial might come to use the ROS install.

>

Now return to the robotpkg folder and install all the set:

<pre><code class="shell">

cd robotpkg

make update-mpcset

</code></pre>

>

During the installation, some required dependencies need to be installed with the usual package manager (e.g. @apt@ on Ubuntu). When the install stops, install the required packages and rerun the above command.

>

h4. 5. Matlab configuration

The last step is to update the Matlab path to use the custom libraries, if relevant.

Add the following paths in the Matlab path window:

<pre><code class="shell">

</path/to/openrobots>/lib/matlab

</path/to/openrobots>/lib/matlab/simulink

</path/to/openrobots>/lib/matlab/simulink/genomix

</code></pre>

(change </path/to/openrobots> to the value of @$ROBOTPKG_BASE@)

h3. II-2. Install custom components

h4. List of the components

The @src/@ folder contains some additional components, in particular:

* *vision-idl*: the type declarations regarding the camera modules

* *camgazebo-genom3*: read the data from the gazebo innate cameras, via the gazebo API

* *camviz-genom3*: record and/or display the images from a camera

* *arucotag-genom3*: detect and the ArUco markers

* *ikf-genom3*: an implementation of the intermittent Kalman filter presented in the paper

* *uavmpc-genom3*: the NMPC controller presented in the paper

* *tagcontrol*: both a Gazebo plugin and a genom component that work together to move the ArUco tags in the Gazebo world

h4. Install the extra components

Since they are not considered 'stable' as the one provided in robotpkg, we rather install them in a devel folder.

Go to the project root, check that the devel folder exists, export the path if you didn't source the @env.sh@. Then go to the sources folder:

<pre><code class="shell">

export DEVEL_BASE=`pwd`/devel

cd src/

</code></pre>

>

For the manual installation, @asciidoctor@ is needed. It can be installed using @apt@ or any package manager.

Each component here has to be installed manually, using @autoconf@. To do so, proceed as follow:

<pre><code class="shell">

cd src/<component>/

./bootstrap.sh

mkdir build

cd build

../configure --prefix=$DEVEL_BASE --with-templates=pocolibs/client/c,pocolibs/server

make install

</code></pre>

>

The component @vision-idl@ has to be installed first since it defines some type headers used by others.

The installation of the main component, @uavmpc-genom3@, is described in the next subsection.

>

h4. Install the MPC controller

Before installing the MPC controller, we have to generate the @C@ sources corresponding to the desired model.

To do so, go to the @model_generation/@ folder:

<pre><code class="shell">

cd src/uavmpc-genom3/model_generation

</code></pre>

>

There is a README.md file there, explaining the requirements.

In short, the model sources are exported to @C@ using @CasADi@ in @python3@.

@python3@ along with @NumPy@, @SciPy@ and @CasADi@  are required, and easily installable on most Linux distributions (e.g. with @apt@ and @pip3@).

Then, the sources are generated using:

<pre><code class="shell">

python3 gen_model.py <quad or hexa>

</code></pre>

Then install the component as explained before, but add the following flags to the @configure@ command:

<pre><code class="shell">

configure --prefix=$GENOM_DEVEL --with-templates=pocolibs/client/c,pocolibs/server CPPFLAGS="-Wall -march=native -mfpmath=sse -I$ROBOTPKG_BASE/include" CXXFLAGS="-O3" CFLAGS="-O3" LDFLAGS="-L$ROBOTPKG_BASE/lib -Wl,-R$ROBOTPKG_BASE/lib"

</code></pre>

>

Inside each model file, the variable @number_agents@ (l. 36) can be changed to specify the number of UAVs in the system. This is needed to define the size of the parameter matrix (containing the other agents' variables).

h3. II-3. Setup the environment

In order to run all the installed executables, we need to set up the path to the newly created folders.

All the required variables are exported in the @env.sh@ file.

h2. III - Running the simulation

The @ws/@ folder contains all the material to run a basic simulation with the NMPC.

In a terminal, launch the @launch_mpc.sh@ script. It takes 1 argument, being the number of UAVs. It starts all the Genom components, in the background. It is used as a console since it displays warnings or errors during runtime.

In another terminal, start gazebo with one of the world files provided.

Finally, run @matlab@ or @eltclsh@ and go to the relevant subsection below.

>

h3. III.1. Running the simulations with Matlab

Change the flag at the top of the script to use either the quadrotor or the hexarotor.

>

The provided scripts are organized as follow

* The @param_*.m@ scripts provide the parameters for either a standard colinear quadrotor (denoted *QR*) or a tilted-propeller hexarotor (denoted *tx*).

* The @init.m@ script that connects all the components together and calls the initialization services for all provided components.

* The @traj_*.m@ that runs the specific trajectories for a specific scenario.

>

Run the init script and wait until it stops displaying in the console.

If no error occurred, run any traj script then press enter between each step to proceed to the next one. The evolution can be watched in gazebo and in the console terminal in parallel.

>

The init.m script contains some parameters at the top of the script, in particular, n_UAV. It will specify the number of UAVs to initialize.

h3. III.2. List of the provided trajectories

The @traj_mpc_1/2/3.m@ files respectively correspond to the number of UAVs to use.

h3. III.3. Comments on how to use the simulator

* The MPC weights and parameters can be tuned in the @param_*.m@ files.

* The marker can be moved, either using a joystick or manually, using the interface of the tagcontrol component.