# Training pipeline

Behavior cloning of analytic herding teachers into a neural network
policy that runs in Webots. PPO from scratch and PPO fine-tune of BC
were tried earlier and are kept under `train_ppo.py` as experimental
options, but the BC route alone is what we ship.

## Files

```
herding_env.py     — Gymnasium env (used for demo collection + eval)
bc_pretrain.py     — supervised MSE+cosine training of an SB3 MlpPolicy
                     against (obs, action) demos
eval.py            — analytic teachers + BC policies, full n=1..10 grid
parity_test.py     — shape/determinism/baseline smoke test
train_ppo.py       — PPO trainer (experimental — see Appendix below)
configs/           — PPO hyperparameter YAML
runs/              — checkpoints (.gitignored)
```

## Setup

```
pip install -r requirements.txt
```

CPU is the default and recommended device — SB3 PPO with an MLP policy
of this size runs faster on CPU than GPU because the bottleneck is
rollout collection, not gradient compute.

## The BC pipeline

```
# 1. Generate demos from an analytic teacher.
#    --teacher: strombom (default), sequential, drive_only, hybrid, strombom_smooth
python -m tools.collect_demos --teacher strombom \
    --out demos.npz --seeds-per-n 30 --subsample 3

# 2. Behavior-clone the demos into an MLP policy.
python -m training.bc_pretrain --demos demos.npz \
    --out runs/bc_flock --epochs 100 --net-arch 512,512

# 3. Evaluate the resulting policy.
python -m training.eval --policy runs/bc_flock \
    --max-flock 10 --max-steps 30000 --n-seeds 5
```

Wall time: ~10 min demos + ~5 min BC training + ~5 min eval.

`bc_pretrain.py` saves the **best-val_cos** snapshot, not the final
epoch — multi-modal teachers (Strömbom's collect/drive switch) make
training noisy and the last epoch is often worse than an earlier one.

## Available analytic teachers

| Name | What it does | Best for |
|---|---|---|
| `strombom` | Canonical Strömbom — collect when flock is scattered, drive CoM otherwise | Tight-cohesion regime, n=1-10 |
| `sequential` | Pick the sheep closest to the pen and drive only it | Loose-cohesion regime, n=1-10 |
| `drive_only` | Strömbom drive without collect mode (continuous action) | Easier-to-BC alternative; less reliable than full Strömbom |
| `hybrid` | Drive rearmost sheep when far, switch to closest near gate | Failed experiment, kept for write-up |
| `strombom_smooth` | Sigmoid-blended Strömbom collect↔drive | Failed experiment |

## Evaluating the analytic teachers directly

```
python -m training.eval --policy strombom    --max-flock 10 --max-steps 30000 --n-seeds 5
python -m training.eval --policy sequential  --max-flock 10 --max-steps 30000 --n-seeds 5
```

## Webots inference

The Webots dog controller (`controllers/shepherd_dog/shepherd_dog.py`)
loads a saved BC zip when launched in `rl` mode:

```
HERDING_POLICY_DIR=$PWD/runs/bc_flock tools/run_webots.sh 10 rl
```

It auto-discovers a checkpoint named `policy.zip`, `best_model.zip`, or
`final.zip` in the directory.

## Appendix — experimental PPO scripts

`train_ppo.py` contains the PPO/RL pipeline tried before BC:
* PPO from scratch with curriculum learning over flock size + spawn area.
* PPO fine-tune of a BC checkpoint.

Both ran into stability issues (PPO's exploration noise destroys BC
weights faster than the reward signal can rebuild them; PPO from
scratch never sees pen events often enough during random exploration to
credit-assign the +500 done bonus).

The script is left in place because the abstractions are sound and the
code is reusable for follow-up work (e.g. KL-regularised fine-tune
with a frozen reference policy). Not part of the deliverable pipeline.