Drop versioning vocabulary, polish docstrings, fix world-aware policy resolution

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    up-to-date — adds `universal`, removes outdated single-policy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>

README.md

@@ -78,13 +78,11 @@ HERDING_USE_GT=1 tools/run_webots.sh 5 strombom differential field
 `make help` lists every target and the overridable hyperparameters.
 
 **Mecanum note**: the `ShepherdDogMecanum.proto` uses physical roller
-hinges in Webots (committed 2026-05-16). The Webots calibration shows
-a ~60% strafe efficiency and ~28% backward bleed compared to textbook
-mecanum; the gym kinematics in `HERDING_MEC_WEBOTS` are tuned to
-match. **Mecanum BC/RL policies need to be retrained against this
-preset** — see `mecanum_proto_gap.md` in `memory/` for the 3-command
-flow. The v1 policies in `training/runs/{bc,rl}_mecanum_*` predate the
-proto rewrite and will not herd reliably in Webots until retrained.
+hinges in Webots. The Webots calibration shows ~60% strafe efficiency
+and ~28% backward bleed compared to textbook mecanum; the gym
+kinematics in `HERDING_MEC_WEBOTS` are tuned to match. **Mecanum BC/RL
+policies need to be retrained against this preset** — see the retrain
+flow in the Mecanum results section below.
 
 ## Documentation map
 
@@ -215,16 +213,30 @@ information.
 
 ### Mecanum (differential is the headline)
 
-The `ShepherdDogMecanum.proto` was rewritten on 2026-05-16 with 32
-physical roller hinges, giving true omnidirectional motion in Webots
-(`tools/calibrate_mecanum.sh` confirms the X-pattern). The mecanum
-calibration shows ~60% strafe efficiency vs textbook (vs ~89% on
-forward), so v1 mecanum BC/RL policies trained on textbook gym
-mecanum no longer herd reliably. The fix is staged but not run:
-the gym now has `HERDING_MEC_WEBOTS` which matches Webots' physical
-mecanum, and `training/bc/collect.py` / `training/rl/train.py` auto-
-select this preset for mecanum runs. Retraining (≈ 2 h per combo,
-4 combos) is the documented future step.
+`ShepherdDogMecanum.proto` has 32 physical roller hinges giving true
+omnidirectional motion in Webots — `tools/calibrate_mecanum.sh`
+confirms the X-pattern. Calibration shows ~60% strafe efficiency vs
+textbook (versus ~89% on forward), so the gym needs to match the
+imperfect physical mecanum for the trained policy to compensate.
+`HERDING_MEC_WEBOTS` is the matched preset; `training/bc/collect.py`
+and `training/rl/train.py` auto-select it for mecanum runs. Mecanum
+policies were trained on the textbook gym, so they need to be
+retrained against `HERDING_MEC_WEBOTS` (≈ 2 h per combo, 4 combos):
+
+```bash
+python -m training.bc.collect \
+    --drive-mode mecanum --world field --use-webots-preset \
+    --out training/bc/demos_mecanum_field.npz
+python -m training.bc.pretrain \
+    --demos training/bc/demos_mecanum_field.npz \
+    --out training/runs/bc_mecanum_field
+python -m training.rl.train \
+    --bc training/runs/bc_mecanum_field \
+    --out training/runs/rl_mecanum_field \
+    --drive-mode mecanum --world field --use-webots-preset
+```
+
+Repeat for `field_round`.
 
 ## License
 

controllers/shepherd_dog/shepherd_dog.py

@@ -1,42 +1,49 @@
 """Shepherd Dog controller (Webots).
 
-Mode is selected by ``HERDING_MODE`` (env var, or via the
-``herding_runtime.cfg`` file the launcher writes since Webots strips
-env vars on some setups):
+Mode is selected by ``HERDING_MODE`` — read from the env var or from
+the ``herding_runtime.cfg`` file the launcher writes (Webots strips
+env vars from controller subprocesses on some setups):
 
     strombom    → canonical Strömbom (2014) collect/drive heuristic
                   wrapped in ActiveScanTeacher (opening rotation +
-                  walk-to-centre when the tracker briefly empties).
-    sequential  → single-target "pin-and-push", same wrapper.
-    bc          → behaviour-cloned MLP, trained on Strömbom demos.
-                  Default policy: training/runs/bc/policy.zip.
-    rl          → KL-regularised PPO fine-tune of bc. Same obs/action
-                  space as bc; refines time-to-pen via reward while
-                  staying anchored to bc.
-                  Default policy: training/runs/rl/policy.zip.
+                  walk-to-centre when the tracker briefly empties)
+    sequential  → single-target "pin-and-push", same wrapper
+    universal   → mecanum-aware teacher (Strömbom + omega + recovery)
+    bc          → behaviour-cloned MLP, trained on universal demos
+    rl          → KL-regularised PPO fine-tune of `bc`
+
+Policy directories are resolved by `policy_loader` from
+``training/runs/{bc,rl}_{drive}_{world}`` with a fallback to
+``training/runs/{bc,rl}`` (legacy single-policy paths).
 
 Sheep perception
 ----------------
-The dog perceives sheep through its **front-mounted 140° LiDAR**
+The dog perceives sheep through its front-mounted 140° LiDAR
 (``protos/ShepherdDog.proto``: 180 rays, 12 m max range). Each step:
 
-    1. Reads ``lidar.getRangeImage()``.
-    2. Runs ``herding.perception.lidar_perception.detections_from_scan``
-       to cluster returns into world-frame ``(x, y)`` sheep estimates.
-    3. Folds those into a ``SheepTracker`` which maintains last-seen
-       positions for sheep currently out of FOV and latches "penned"
-       once a track crosses the gate plane south.
+    1. Read ``lidar.getRangeImage()``.
+    2. Cluster returns into world-frame ``(x, y)`` estimates
+       (``herding.perception.lidar_perception.detections_from_scan``).
+    3. Fold detections into a ``SheepTracker``, which maintains
+       last-seen positions for sheep currently out of FOV, requires
+       consensus across multiple frames before promoting a candidate
+       to a real track, and latches "penned" once a track crosses
+       the gate plane south.
 
-Sheep ``emitter`` messages are read **for diagnostic logging only**
-(GT_penned counter + auto-finish sentinel); they are never used to
-drive the policy. Perception for control comes entirely from LiDAR.
+Setting ``HERDING_USE_GT=1`` bypasses the tracker and feeds emitter
+ground-truth positions to the policy — useful as a perception
+ablation for the analytic baselines.
+
+Sheep emitter messages are otherwise read for diagnostic logging
+only (``GT_penned`` counter + auto-finish sentinel); the control
+loop never depends on them.
 
 Auto-finish
 -----------
-When the dog observes (via GT, read off the receiver) that all sheep
-are penned, it writes ``training/.run_done`` and the launcher
-(``tools/run_webots.sh``) detects it and closes Webots. This keeps
-batch evaluation runs bounded.
+When every emitter-reported sheep is penned, the controller writes
+``training/.run_done``. The launcher (``tools/run_webots.sh``)
+detects the sentinel and closes Webots so headless sweep runs are
+bounded.
 """
 
 import math
@@ -111,6 +118,24 @@ MODE = (os.environ.get("HERDING_MODE")
         or _runtime_cfg.get("HERDING_MODE")
         or "bc").lower()
 
+_VALID_MODES = ("bc", "rl", "strombom", "sequential", "universal", "calibrate")
+if MODE not in _VALID_MODES:
+    print(f"[dog] unknown HERDING_MODE={MODE!r}; defaulting to strombom.")
+    MODE = "strombom"
+
+# Drive mode: "differential" (2-wheel) or "mecanum" (4-wheel omnidirectional).
+DRIVE_MODE = (os.environ.get("HERDING_DRIVE")
+              or _runtime_cfg.get("HERDING_DRIVE")
+              or "differential").lower()
+if DRIVE_MODE not in ("differential", "mecanum"):
+    print(f"[dog] unknown HERDING_DRIVE={DRIVE_MODE!r}; defaulting to differential.")
+    DRIVE_MODE = "differential"
+
+# World shape — used to disambiguate the trained policy directory.
+WORLD = (os.environ.get("HERDING_WORLD")
+         or _runtime_cfg.get("HERDING_WORLD")
+         or "field").lower()
+
 # Diagnostic: bypass LiDAR tracker and use GT emitter positions directly.
 # Set HERDING_USE_GT=1 to isolate perception sim-to-real gap from policy quality.
 USE_GT_PERCEPTION = bool(int(
@@ -119,50 +144,34 @@ USE_GT_PERCEPTION = bool(int(
 ))
 
 
-def _resolve_policy_dir(mode: str) -> str:
-    """Where to look for the trained policy for the given mode.
+def _resolve_policy_dir(mode: str, drive: str, world: str) -> str:
+    """Where to look for the trained policy for the given mode/drive/world.
 
     Priority:
       1. HERDING_POLICY_DIR env var or runtime-cfg entry, if it points
          to a real directory.
-      2. Drive-mode-specific default:
-            bc → training/runs/bc_differential (or bc_mecanum)
-            rl → training/runs/rl_differential (or rl_mecanum)
-      3. Legacy path (no drive suffix):
-            bc → training/runs/bc
-            rl → training/runs/rl
+      2. Canonical:  training/runs/{bc,rl}_<drive>_<world>
+      3. Drive-only: training/runs/{bc,rl}_<drive>
+      4. Bare-mode:  training/runs/{bc,rl}
+    The first existing directory wins; if none exist, the canonical
+    path is returned so the loader's error message is informative.
     """
     env_dir = (os.environ.get("HERDING_POLICY_DIR")
                or _runtime_cfg.get("HERDING_POLICY_DIR"))
     if env_dir and os.path.isdir(env_dir):
         return env_dir
-    drive = DRIVE_MODE
-    mode_default = {
-        "bc": os.path.join(_PROJECT_ROOT, "training", "runs",
-                           f"bc_{drive}"),
-        "rl": os.path.join(_PROJECT_ROOT, "training", "runs",
-                           f"rl_{drive}"),
-    }
-    primary = mode_default.get(mode, mode_default["bc"])
-    if os.path.isdir(primary):
-        return primary
-    # Fallback: legacy paths without drive suffix.
-    legacy = {
-        "bc": os.path.join(_PROJECT_ROOT, "training", "runs", "bc"),
-        "rl": os.path.join(_PROJECT_ROOT, "training", "runs", "rl"),
-    }
-    fallback = legacy.get(mode, legacy["bc"])
-    if os.path.isdir(fallback):
-        return fallback
-    return env_dir or primary
+    base = "rl" if mode == "rl" else "bc"
+    runs = os.path.join(_PROJECT_ROOT, "training", "runs")
+    for cand in (f"{base}_{drive}_{world}", f"{base}_{drive}", base):
+        path = os.path.join(runs, cand)
+        if os.path.isdir(path):
+            return path
+    return os.path.join(runs, f"{base}_{drive}_{world}")
 
 
-_VALID_MODES = ("bc", "rl", "strombom", "sequential", "universal", "calibrate")
-if MODE not in _VALID_MODES:
-    print(f"[dog] unknown HERDING_MODE={MODE!r}; defaulting to strombom.")
-    MODE = "strombom"
+print(f"[dog] mode={MODE}  drive={DRIVE_MODE}  world={WORLD}")
 
-POLICY_DIR = _resolve_policy_dir(MODE)
+POLICY_DIR = _resolve_policy_dir(MODE, DRIVE_MODE, WORLD)
 policy_handle = None
 if MODE in ("bc", "rl"):
     print(f"[dog] resolved POLICY_DIR={POLICY_DIR}  exists={os.path.isdir(POLICY_DIR)}")
@@ -173,16 +182,6 @@ if MODE in ("bc", "rl"):
     except Exception as exc:
         print(f"[dog] policy load failed ({exc!r}); falling back to strombom.")
         MODE = "strombom"
-print(f"[dog] running in mode={MODE}")
-
-# Drive mode: "differential" (2-wheel) or "mecanum" (4-wheel omnidirectional).
-DRIVE_MODE = (os.environ.get("HERDING_DRIVE")
-              or _runtime_cfg.get("HERDING_DRIVE")
-              or "differential").lower()
-if DRIVE_MODE not in ("differential", "mecanum"):
-    print(f"[dog] unknown HERDING_DRIVE={DRIVE_MODE!r}; defaulting to differential.")
-    DRIVE_MODE = "differential"
-print(f"[dog] drive mode={DRIVE_MODE}")
 
 
 # ---------------------------------------------------------------------------

herding/perception/sheep_tracker.py

@@ -1,32 +1,26 @@
 """Multi-target tracker for LiDAR-detected sheep.
 
-Greedy nearest-neighbour data association across frames, with a wider
-re-acquisition gate for stale tracks (sheep flee during occlusion and
-reappear off-position), plus memory of last-seen positions for sheep
-out of FOV. Output is ``{name: (x, y)}`` — Strömbom / Sequential
-consume it directly.
+Three-stage greedy nearest-neighbour data association:
 
-When **predictive mode** is enabled (the default), tracks carry a
-constant-velocity state ``(vx, vy)`` estimated from the last two
-observations. While a track is occluded its position is extrapolated
-using this velocity for up to ``PREDICT_STEPS`` frames, keeping the
-teacher's CoM estimate stable during brief losses. After prediction
-expires, the track falls back to its last-seen position (static memory)
-until ``FORGET_STEPS`` deletes it entirely.
+1. **Consensus promotion**. New detections start as *candidate* tracks
+   invisible to ``get_positions``. They must accumulate ``consensus_k``
+   matches within ``consensus_radius_m`` to promote; candidates that
+   fail to re-confirm within ``consensus_max_age`` steps die. This
+   filters one-shot LiDAR phantoms — wall returns, multi-cluster sheep
+   splits, fast-moving sheep position jumps — at the cost of a small
+   acquisition latency (~50 ms at the default ``consensus_k=3``).
+   ``consensus_k=1`` disables the stage.
+2. **Constant-velocity prediction**. Each track carries a smoothed
+   ``(vx, vy)``. While a track is occluded its position is
+   extrapolated for up to ``PREDICT_STEPS`` frames, then falls back to
+   last-seen static memory until ``FORGET_STEPS`` deletes it.
+3. **Pen latching**. A track whose estimated position crosses the gate
+   plane south of ``is_penned_position`` is marked penned, excluded
+   from ``get_positions``, and kept indefinitely.
 
-A track is marked penned once its estimated position crosses the gate
-plane south (``is_penned_position``). Penned tracks are excluded from
-``get_positions`` and kept indefinitely.
-
-**Consensus promotion** (``consensus_k > 1``): every new detection
-starts as a *candidate* track that is invisible to ``get_positions``.
-It must be matched ``consensus_k`` times within a tight radius
-(``consensus_radius_m``) before being promoted to a regular track.
-Candidates that fail to re-confirm within ``consensus_max_age`` steps
-are deleted. The cost is a small acquisition latency
-(``consensus_k * timestep`` ≈ 65 ms) in exchange for rejecting the
-one-shot LiDAR phantom returns Webots produces from real-world 3D
-geometry. ``consensus_k=1`` disables the stage entirely (default).
+Output of :meth:`SheepTracker.get_positions` is ``{name: (x, y)}`` —
+Strömbom, Sequential and the BC observation builder consume it
+directly.
 """
 
 from __future__ import annotations

tools/run_webots.sh

@@ -60,19 +60,26 @@ DST="$ROOT/worlds/${WORLD}_test.wbt"
 if [[ -n "${HERDING_POLICY_DIR:-}" ]]; then
     RESOLVED_POLICY_DIR="$HERDING_POLICY_DIR"
 else
-    # Try drive-mode-specific path first, then legacy path.
+    # The training pipeline writes policies to:
+    #   training/runs/{bc,rl}_<drive>_<world>
+    # Try that first; fall back to the drive-only and finally the
+    # bare-mode legacy paths so older policy checkouts still load.
     if [[ "$MODE" == "rl" ]]; then
-        DRIVED="$ROOT/training/runs/rl_${DRIVE}"
-        LEGACY="$ROOT/training/runs/rl"
+        BASE="rl"
     else
-        DRIVED="$ROOT/training/runs/bc_${DRIVE}"
-        LEGACY="$ROOT/training/runs/bc"
+        BASE="bc"
     fi
-    if [[ -d "$DRIVED" ]]; then
-        RESOLVED_POLICY_DIR="$DRIVED"
-    else
-        RESOLVED_POLICY_DIR="$LEGACY"
+    for CAND in \
+        "$ROOT/training/runs/${BASE}_${DRIVE}_${WORLD}" \
+        "$ROOT/training/runs/${BASE}_${DRIVE}" \
+        "$ROOT/training/runs/${BASE}"
+    do
+        if [[ -d "$CAND" ]]; then
+            RESOLVED_POLICY_DIR="$CAND"
+            break
         fi
+    done
+    : "${RESOLVED_POLICY_DIR:=$ROOT/training/runs/${BASE}_${DRIVE}_${WORLD}}"
 fi
 
 cp "$SRC" "$DST"

training/README.md

@@ -1,80 +1,105 @@
-# Training and Evaluation Details
+# Training and evaluation details
 
-This file is the command-level companion to the root README. It focuses
-on data collection, BC, PPO fine-tuning, evaluation flags, and generated
-artifacts; use the root README for the high-level architecture and
-Webots demo quick start.
+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.
 
-Two stages, strictly sequential:
+The pipeline is two strictly-sequential stages per `(drive, world)`
+combo:
 
 ```
-sim demos (Strömbom on tracker output, K=4 frame stack)
+sim demos (universal teacher on tracker output, K=4 frame stack)
     │
     ▼
-bc/pretrain.py  ──►  runs/bc   (Strömbom-imitated MLP)
+bc/pretrain.py  ──►  runs/bc_<drive>_<world>   (MLP)
     │
     ▼  KL-regularised PPO fine-tune
     │
-runs/rl                        (deployed `rl` mode — beats BC and Strömbom)
+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 (whitelisted entries in top-level .gitignore)
-
-(Unit + integration tests live in the top-level ``tests/`` directory;
-run with ``python -m pytest tests/``.)
+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 run everything is the Makefile at the project
-root: ``make`` does the full chain, ``make rl`` rebuilds whatever's
-needed up to that point, etc. The individual stages below are kept
-explicit for cases where you want to tune a single step.
+The simplest way to train one combo is the project-root Makefile:
 
 ```bash
-# 1. Sim demos with the active-scan + Strömbom teacher under LiDAR
-#    perception. K=4 frame stack so the MLP has temporal context.
-python -m training.bc.collect --teacher strombom \
-    --out training/bc/demos.npz --seeds-per-n 15 --subsample 3 --frame-stack 4
+make DRIVE=differential WORLD=field           # demos → bc → rl → eval
+make DRIVE=differential WORLD=field_round
+make train_all                                # all four combos sequentially
+```
 
-# 2. Behaviour-clone.
-python -m training.bc.pretrain --demos training/bc/demos.npz \
-    --out training/runs/bc --epochs 60 --net-arch 512,512
+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 --out training/runs/rl \
+    --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 \
+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.
+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.
 
-## Available analytic teachers
+## 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 | Default; works for n=1–10 under tight cohesion |
-| `sequential` | Pick the sheep closest to the pen and drive only it | Alternative; needs loose-cohesion regime |
+| `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 |
 
-Both are wrapped at demo-collection time in
+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
@@ -83,8 +108,11 @@ 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
 ```
-python -m training.eval --policy strombom    --max-flock 10 --max-steps 15000 --n-seeds 10
-python -m training.eval --policy sequential  --max-flock 10 --max-steps 15000 --n-seeds 10
-```
-

training/bc/collect.py

@@ -8,8 +8,8 @@ the same partial-obs view the student will have at deployment.
 
 Usage::
 
-    python -m training.bc.collect --teacher strombom \\
-        --out training/bc/demos.npz --frame-stack 4
+    python -m training.bc.collect --teacher universal --drive-mode differential \\
+        --world field --out training/bc/demos_differential_field.npz --frame-stack 4
 """
 
 from __future__ import annotations
@@ -125,7 +125,9 @@ def collect_one(n_sheep: int, seed: int, max_steps: int, subsample: int,
 
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--out", default="training/bc/demos.npz")
+    parser.add_argument("--out", required=True,
+                        help="Output .npz path (convention: "
+                             "training/bc/demos_<drive>_<world>.npz).")
     parser.add_argument("--n-sheep-list", default="1,2,3,5,8,10")
     parser.add_argument("--seeds-per-n", type=int, default=15)
     parser.add_argument("--max-steps", type=int, default=30000)

training/bc/pretrain.py

@@ -12,8 +12,8 @@ Output zip is loadable by ``PPO.load(...)`` and consumed by
 Usage::
 
     python -m training.bc.pretrain \\
-        --demos training/bc/demos.npz \\
-        --out training/runs/bc
+        --demos training/bc/demos_differential_field.npz \\
+        --out training/runs/bc_differential_field
 """
 
 from __future__ import annotations
@@ -70,8 +70,11 @@ def policy_forward_mean(policy, obs_batch):
 
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--demos", default="training/bc/demos.npz")
-    parser.add_argument("--out", default="training/runs/bc")
+    parser.add_argument("--demos", required=True,
+                        help="Path to demos .npz collected by training.bc.collect.")
+    parser.add_argument("--out", required=True,
+                        help="Output directory (convention: "
+                             "training/runs/bc_<drive>_<world>).")
     parser.add_argument("--epochs", type=int, default=60)
     parser.add_argument("--batch-size", type=int, default=256)
     parser.add_argument("--lr", type=float, default=1e-3)