Johnny Fernandes
10c01a938e
User-facing pass after the project was decided to be a single
submission with no inner iterations.
* Remove every "v1"/"v2"/"versioning" reference from the docs:
- README mecanum section trims the "v1 predates the rewrite" prose
in favour of a self-contained retrain recipe.
- The 3.2 GB `training/runs/v1_clean/` backup directory is deleted.
* Refresh control-layer docstrings:
- `sheep_tracker.py` header now describes the three actual pipeline
stages (consensus, prediction, pen latching) instead of layering
the consensus stage on top of a stale "predictive mode" preamble.
- `controllers/shepherd_dog/shepherd_dog.py` mode list is
up-to-date — adds `universal`, removes outdated single-policy
default paths, mentions `HERDING_USE_GT=1` as the perception
ablation.
* Refresh training command examples:
- `training/bc/collect.py` and `training/bc/pretrain.py` usage
snippets show the world-suffixed paths the Makefile actually
uses; the `--out` arg is now required so old "demos.npz"
invocations error loudly instead of silently overwriting.
- `training/README.md` rewritten — drops the legacy `runs/bc`
diagram, documents the per-(drive, world) pipeline, and adds
the mecanum retraining caveat.
* Fix policy-directory resolution end-to-end:
- `tools/run_webots.sh` now tries
`training/runs/{bc,rl}_<drive>_<world>` first, then the drive-
only path, then the bare-mode legacy path — matching the actual
on-disk layout. Previously it looked for `bc_<drive>` (no
world) and silently fell back to `bc`, masking the world
selection.
- `controllers/shepherd_dog/shepherd_dog.py:_resolve_policy_dir`
has the same fix plus a latent NameError unmasked: it referenced
`DRIVE_MODE` before that variable was set at module load. The
block is restructured so MODE/DRIVE_MODE/WORLD are resolved
first, then the function uses them as explicit arguments.
126 pytest cases still pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
119 lines
4.4 KiB
Markdown
119 lines
4.4 KiB
Markdown
# Training and evaluation details
|
||
|
||
Command-level companion to the root README. Covers demo collection,
|
||
behaviour cloning, PPO fine-tuning, and evaluation flags; use the root
|
||
README for the high-level architecture and Webots quick start.
|
||
|
||
The pipeline is two strictly-sequential stages per `(drive, world)`
|
||
combo:
|
||
|
||
```
|
||
sim demos (universal teacher on tracker output, K=4 frame stack)
|
||
│
|
||
▼
|
||
bc/pretrain.py ──► runs/bc_<drive>_<world> (MLP)
|
||
│
|
||
▼ KL-regularised PPO fine-tune
|
||
│
|
||
runs/rl_<drive>_<world> (deployed `rl` mode)
|
||
```
|
||
|
||
## Files
|
||
|
||
```
|
||
herding_env.py — Gymnasium env (LiDAR raycast + tracker by default)
|
||
bc/collect.py — universal-teacher sim demos
|
||
bc/pretrain.py — MSE + cosine BC of (obs, action) demos into MlpPolicy
|
||
rl/train.py — KL-regularised PPO fine-tune of a BC checkpoint
|
||
rl/train_lstm.py — RecurrentPPO variant (ablation)
|
||
eval.py — multi-seed analytic / learned policy comparison
|
||
runs/ — checkpoints (gitignored except for policy.zip)
|
||
```
|
||
|
||
Unit + integration tests live in the top-level `tests/`. Run with
|
||
`make test` or `python -m pytest tests/`.
|
||
|
||
## End-to-end pipeline
|
||
|
||
The simplest way to train one combo is the project-root Makefile:
|
||
|
||
```bash
|
||
make DRIVE=differential WORLD=field # demos → bc → rl → eval
|
||
make DRIVE=differential WORLD=field_round
|
||
make train_all # all four combos sequentially
|
||
```
|
||
|
||
The individual stages below are kept explicit for cases where you
|
||
want to tune a single step.
|
||
|
||
```bash
|
||
# 1. Sim demos with the active-scan + universal teacher under LiDAR
|
||
# perception. K=4 frame stack so the MLP has temporal context.
|
||
python -m training.bc.collect \
|
||
--teacher universal --drive-mode differential --world field \
|
||
--out training/bc/demos_differential_field.npz \
|
||
--seeds-per-n 15 --subsample 3 --frame-stack 4
|
||
|
||
# 2. Behaviour-clone the demos.
|
||
python -m training.bc.pretrain \
|
||
--demos training/bc/demos_differential_field.npz \
|
||
--out training/runs/bc_differential_field \
|
||
--epochs 60 --net-arch 512,512
|
||
|
||
# 3. KL-regularised PPO fine-tune of bc.
|
||
python -m training.rl.train \
|
||
--bc training/runs/bc_differential_field \
|
||
--out training/runs/rl_differential_field \
|
||
--drive-mode differential --world field \
|
||
--total-timesteps 1000000
|
||
|
||
# 4. Multi-seed eval (env-side, fast).
|
||
python -m training.eval --policy training/runs/rl_differential_field \
|
||
--drive-mode differential --world field \
|
||
--max-flock 10 --max-steps 15000 --n-seeds 10
|
||
```
|
||
|
||
`bc/pretrain.py` saves the **best-val_cos** snapshot, not the final
|
||
epoch — multi-modal teachers make training noisy and the last epoch
|
||
is often worse than an earlier one.
|
||
|
||
`rl/train.py` loads BC weights into both a trainable policy and a
|
||
frozen reference, fixes `log_std` small, and adds `β · KL(π‖π_ref)` to
|
||
the loss so the policy can only move within a trust region around BC.
|
||
See the file header for hyperparameter rationale.
|
||
|
||
## Mecanum retraining
|
||
|
||
For mecanum runs, pass `--use-webots-preset`. Both `collect.py` and
|
||
`train.py` detect `--drive-mode mecanum` and switch to the
|
||
`HERDING_MEC_WEBOTS` preset, which matches the physical-roller
|
||
Webots proto's strafe efficiency (~0.4) and forward bleed (~−0.28).
|
||
Training without this preset produces a policy that herds in textbook
|
||
gym mecanum but not in Webots.
|
||
|
||
## Analytic teachers
|
||
|
||
| Name | What it does | Notes |
|
||
|---|---|---|
|
||
| `strombom` | Strömbom 2014 — collect when flock is scattered, drive CoM otherwise | Round-world aware (radially-inward fallback when natural target lies outside the curved boundary) |
|
||
| `sequential` | Three-phase: collect, drive, then single-target push for the last 1–2 stragglers | Alternative to strombom |
|
||
| `universal` | Strömbom core + mecanum omega + last-straggler recovery | Used as the BC demo teacher |
|
||
|
||
All three are wrapped at demo-collection time in
|
||
`herding/control/active_scan.py:ActiveScanTeacher`, which adds an
|
||
opening in-place rotation, walk-to-centre when the LiDAR sees
|
||
nothing, and near-sheep speed modulation (same modulation
|
||
`herding/control/modulation.py` applies to every dog mode at
|
||
inference).
|
||
|
||
## Evaluating analytic teachers directly
|
||
|
||
```bash
|
||
python -m training.eval --policy strombom \
|
||
--drive-mode differential --world field \
|
||
--max-flock 10 --max-steps 15000 --n-seeds 10
|
||
python -m training.eval --policy sequential \
|
||
--drive-mode differential --world field_round \
|
||
--max-flock 10 --max-steps 15000 --n-seeds 10
|
||
```
|