Approach v3 w/ south penalty

controllers/shepherd_dog_rl/shepherd_dog_rl.py

@@ -59,6 +59,13 @@ VECNORM_PATH = os.path.join(_HERE, "vecnorm.pkl")
 DEBUG_CSV    = os.path.join(_HERE, "debug.csv")
 DEBUG_ENABLED = True   # set False to disable debug.csv logging
 
+# ── action smoothing ─────────────────────────────────────────────────────────
+# EMA on policy output to suppress the rapid oscillation (vx/vy flipping
+# between -1 and +1 every step) that stalls the physical dog.  0 = no
+# smoothing (raw policy), 1 = frozen.  0.3 keeps ~30% of previous action.
+ACTION_SMOOTH = 0.3
+prev_action   = np.zeros(2, dtype=np.float32)
+
 
 def norm_angle(a: float) -> float:
     while a >  math.pi: a -= 2 * math.pi
@@ -227,9 +234,15 @@ while robot.step(timestep) != -1:
     raw_obs  = build_obs(dog_pos, sheep_positions, n_sheep)
     obs_norm = vecnorm.normalize_obs(raw_obs[np.newaxis])  # (1, 13)
 
-    # 4. Policy inference
+    # 4. Policy inference + smoothing
     action, _ = model.predict(obs_norm, deterministic=True)
-    vx, vy    = float(action[0][0]), float(action[0][1])
+    raw_a = np.array([float(action[0][0]), float(action[0][1])], dtype=np.float32)
+    if ACTION_SMOOTH > 0:
+        smoothed = ACTION_SMOOTH * prev_action + (1.0 - ACTION_SMOOTH) * raw_a
+        prev_action[:] = smoothed
+        vx, vy = float(smoothed[0]), float(smoothed[1])
+    else:
+        vx, vy = float(raw_a[0]), float(raw_a[1])
 
     # 5. Drive
     drive(vx, vy)

training/herding_env.py

@@ -60,6 +60,11 @@ class HerdingEnv(gym.Env):
     W_PEN_BONUS = 10.0   # per sheep penned
     W_COMPLETE  = 100.0  # all sheep penned
     W_STEP_COST = 0.02   # time penalty — strong enough to punish doing nothing
+    W_SOUTH     = 0.01   # per-sheep per-metre penalty for active sheep below the pen
+                         # entrance (y < PEN_Y[1]=-8). Keeps the dog from letting
+                         # sheep drift into the dead zone below the open face where
+                         # they must reverse direction (north) to enter — hard to
+                         # recover. 0.01 ≈ half step_cost per metre below per sheep.
     W_COMPACT   = 0.0    # reward for flock-radius reduction (off by default)
     W_WALL_TOUCH = 0.01  # per-sheep max penalty at wall surface. Linear ramp
                          # within WALL_TOUCH_BUFFER. Covers field outer walls and
@@ -437,6 +442,16 @@ class HerdingEnv(gym.Env):
         else:
             r_wall_touch = 0.0
 
+        # South penalty: discourage active sheep from drifting below the pen
+        # entrance (y < PEN_Y[1]). Sheep in this zone must reverse direction
+        # (move north) to enter — very hard for the dog to recover from.
+        if self.W_SOUTH and active.any():
+            pts = self.sheep_pos[:self.n_sheep][active]
+            depth = np.maximum(0.0, self.PEN_Y[1] - pts[:, 1])  # metres below entrance
+            r_south = -float(depth.sum()) * self.W_SOUTH
+        else:
+            r_south = 0.0
+
         # Compactness shaping: reward decreases in flock radius (active sheep only)
         if self.W_COMPACT and active.any():
             cur_radius = float(np.linalg.norm(
@@ -450,11 +465,12 @@ class HerdingEnv(gym.Env):
         r_pen_bonus  = newly_penned * self.W_PEN_BONUS
         r_step_cost  = -self.W_STEP_COST
         r_complete   = self.W_COMPLETE if n_penned == self.n_sheep else 0.0
-        reward = (r_progress + alignment + r_compact + r_wall_touch
+        reward = (r_progress + alignment + r_south + r_compact + r_wall_touch
                   + r_pen_bonus + r_step_cost + r_complete)
         rcomps = {
             "progress":   float(r_progress),
             "alignment":  float(alignment),
+            "south":      float(r_south),
             "compact":    float(r_compact),
             "wall_touch": float(r_wall_touch),
             "pen_bonus":  float(r_pen_bonus),

training/train.py

@@ -200,6 +200,7 @@ DEFAULT_CONFIG = {
     "W_PEN_BONUS": 10.0,
     "W_COMPLETE": 100.0,
     "W_STEP_COST": 0.02,
+    "W_SOUTH": 0.01,
     "W_COMPACT": 0.0,
     "W_WALL_TOUCH": 0.04,
     "WALL_TOUCH_BUFFER": 0.3,