Configuration

DroneWatch uses a two-stage configuration flow:

1. OmegaConf composes YAML fragments selected by the root config file and its defaults list.
2. Pydantic validates the resolved result against typed models in src/dronewatch/config/schema.py.

This means the YAML structure is flexible enough for grouped presets, while the final config still has a strict schema.

Root Config Files

The repository currently uses different root configs for different workflows:

• configs/config.yaml for normal PPO training
• configs/evaluate.yaml for standalone checkpoint evaluation
• configs/random_policy.yaml for the random policy baseline
• configs/tune_ppo.yaml for Ray Tune sweeps

Group Config Files

Group config files define parameters of a configuration subgroup, e.g. for logging or network definitions.

• configs/env/ for environment and reward weight definitions
• configs/evaluation/ for evaluation specifications
• configs/logging/ for logging (mlflow) specifications
• configs/model/ for specifying the network of the trained agent (feedforward or lstm)
• configs/random_policy/ for specifying the random policy rollout
• configs/rendering/ for rendering/ gif creation specifications
• configs/training/ for ray rllib training configurations
• configs/tune/ for ray rllib tune configuarions

Group Overrides vs Field Overrides

You can specify which root config to use, overwrite the group configs or overwrite singular fields of the group configs.

Group override

A group override replaces one preset from a config group.

uv run python -m dronewatch.training.train_ppo \
  --config configs/config.yaml \
  model=ppo_feedforward

Here model=ppo_feedforward switches the selected model file that will be used under configs/model/ to the feedforward one.

Field override

A field override changes a specific value inside the composed config.

uv run python -m dronewatch.training.train_ppo \
  --config configs/config.yaml \
  training.stop.iterations=1 \
  training.ppo.minibatch_size=64

Here the selected config groups stay the same, but the final values (stop.iterations and ppo.minibatch_size) change before validation.

How Composition Works

The loader functions in src/dronewatch/config/loader.py all follow the same pattern:

1. Load the root config
2. Read the defaults list
3. Resolve each config group from configs/<group>/<name>.yaml
4. Merge the selected group files
5. Apply CLI key=value overrides
6. Resolve interpolations
7. Validate the final payload with Pydantic

The main entrypoints are:

• load_config() for PPO training
• load_tune_config() for Tune
• load_evaluation_config() for checkpoint evaluation
• load_random_policy_config() for the random baseline

Resolved Configs

Each run writes a resolved YAML snapshot named resolved_config.yaml beside the artifact directory for that run. This is the best source of truth when you need to know what actually ran after group selection and CLI overrides were applied.