Project Balloon (February 2021 - April 2022)

This project is part of a semester- and master-thesis with the goal to get a balloon from a starting point to a pre-defined target point in 3D-space.

Installation

Create a new conda environment

cd project_balloon
conda create --name balloon --file requirements.txt
conda activate balloon
conda install pip
pip install descartes==1.1.0
pip install gym==0.22.0

Get torch depending on your hardware

Getting started

Folder structure

• arduino
  • Was used to control the indoor prototype during the tests in London
• python3
  • Contains everything regarding learning

cd python3

Training (requires downloading a training set)

Uncomment the line import agent_train in setup.py.
It's recommended to generate multiple config files (see generate_yaml.py) and train in parallel.

python3 setup.py yaml/config_train.yaml

Testing with pretrained model

The data provided in the data_example folder is a collection of eleven wind maps used during real life flight tests.

python3 setup.py yaml/config_test.yaml

Visualization

Can be turned on or off in config_NNNNN.py through the parameter render.

• Top window: x-plane (side view)
• Middle window: y-plane (side view)
• Bottom window: z-plane (top-down view)
• White point: Target with acceptance radius
• Red point: balloon
• Red line: current action predicted by algorithm
• White line: path taken by balloon
• Grey line: random roll out used to set reachable target
• Orange background: wind blowing from left to right
• Blue background: wind blowing from right to left
  
  

Advanced steps

Generate config files for data acquisition or training.

Multiple files are generated to allow data acquisition or training in parallel (e.g. on a cluster). If you're running this on a standard computer we suggest only training with a few config files at the same time.
If you are using a cluster, adapt and use the generated submit.txt file to submit jobs easier. The following command will store files in the yaml folder.

python3 generate_yaml.py

Wind Data acquisition

There are two modes of wind-data acquisition. Set big_file parameter in convert_meteomatics.py accordingly.
Either a single dataset with wind data of the current day is acquired, or a big file with data spanning over a month is downloaded which will then have to be further cut into peaces by build_set.py.
Daily wind data set:

cd meteomatics
python3 meteomatics/convert_meteomatics.py ../yaml/config_NNNNN.py

For a big set over multiple months we suggest that you generate one config file for each month (parameter m in config file) and download them in parallel.
In our experiments we used data from a full year (2021). Our training set consists of 1000 wind maps, the testing set of 300 wind maps.
:

cd meteomatics
python3 meteomatics/convert_meteomatics.py ../yaml/config_NNNN0.py
python3 meteomatics/convert_meteomatics.py ../yaml/config_NNNN1.py
...
python3 meteomatics/convert_meteomatics.py ../yaml/config_NNN12.py
python3 build_set.py ../yaml/config_NNNN0.py

python3 generate_noise.py yaml/config_NNNNN.py

Generate noise

This generates noise files (folder noise_MAPSIZE_XxMAPSIZE_Y) that are needed for training and testing.

python3 generate_noise.py