Getting Started¶

Prerequisites¶

DroneWatch currently assumes:

Python 3.11 or newer
uv for package management
Rust and Cargo
Docker only if you want the validation image or MLflow service through Compose

Install the Project¶

The main local setup command is:

make install

This performs the python and rust project installation:

uv sync --dev
uv run maturin develop -m rust/swarm_sim/Cargo.toml

Main Workflows¶

Most commands can be run efficiently via a makefile.
See all available commands with make help

Train PPO¶

make train-ppo

The training prints the results of each training iteration to the command line and logs out to mlflow.
Periodic evaluations are run during training.
Artifacts of these evaluations (checkpoints, gifs, reports) are stored in /artifacts/.
If not configured otherwise, mlflow logs are stored in /outputs/mlruns/.

Evaluate a checkpoint¶

make evaluate-ppo CHECKPOINT=artifacts/checkpoints/ppo/path-to-checkpoint

Similar to the training script, logging and evaluations outputs are stored in /outputs/mlruns/ and /artifacts.

Run Ray Tune search¶

make tune-ppo

Similar to the training script, logging and evaluations outputs are stored in /outputs/mlruns/ and /artifacts.

View MLflow runs¶

make mlflow-up

Run tests¶

make test

Run a baseline random policy¶

make rollout-random

Build and Run with Docker¶

The repository includes a Docker image that starts PPO training directly.

Build it from the repository root:

docker build -t dronewatch:train-entrypoint .

The image copies the repository into /app and uses this entrypoint:

uv run python -m dronewatch.training.train_ppo

That means arguments added after the image name in docker run are passed straight to the training script.

Run the default training config¶

docker run --rm dronewatch:train-entrypoint

Choose a different experiment file¶

docker run --rm dronewatch:train-entrypoint --config configs/debug.yaml

Pass additional config overrides¶

docker run --rm dronewatch:train-entrypoint \
    --config configs/config.yaml \
    training.stop.iterations=10 \
    model=ppo_feedforward

Persist checkpoints, reports, and MLflow outputs on the host¶

By default, container-local outputs disappear when the container is removed. Mount the output directories if you want to keep them:

docker run --rm \
    -v "$(pwd)/artifacts:/app/artifacts" \
    -v "$(pwd)/outputs:/app/outputs" \
    dronewatch:train-entrypoint \
    --config configs/debug.yaml

This keeps checkpoint, report, GIF, and MLflow output files in the local repository directories.