Robust Swarm

A reinforcement learning-based agent that dynamically reallocates tasks and optimizes paths to keep drone swarms mission-cohesive, even amid electronic warfare and harsh weather.

We train the drone swarm to send multiple live streams back to the tower by means of a relay drone. Half way through we take out the communication of half of the drones to simulate a jamming attack. The swarm learns to reallocate the tasks to keep the livestream up.

Watch Video

Environment

Actions and Observations

Reward Function

For each agent $i$ with position $p_i$, the reward $R_i$ is:

$$ \begin{equation} R_i = \begin{cases} e^{-\frac{|d_t - d_{opt}|}{5}} \cdot e^{-\frac{|p_r - \frac{p_i}{2}|}{5}} & \text{if streaming} \\ -1 & \text{if } |p_i - p_j| < 2r_c \text{ for any agent } j \neq i \end{cases} \end{equation} $$

where:

• $d_t = |p_i - p_t|$ is the distance to target
• $d_{opt}$ is the optimal target distance
• $p_r$ is the receiver position
• $r_c$ is the collision radius

Running Instructions

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.in
python src/train.py --help
python src/replay.py --help

If the gpu cannot be found on, you may need to reinstall torch with the right version of cuda for example if you have cuda 12.1 installed, you can run:

pip install --upgrade --force-reinstall torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 --user