Sheep training flock _ improver

training/herding_env.py

@@ -335,7 +335,8 @@ class HerdingEnv(gym.Env):
 
     def _step_sheep(self, i: int) -> np.ndarray:
         """Apply one timestep of boid dynamics to sheep i (mirrors sheep.py)."""
-        pos = self.sheep_pos[i].copy()
+        old_pos = self.sheep_pos[i].copy()   # saved for pen wall collision check
+        pos     = old_pos.copy()
         fx, fy = 0.0, 0.0
         fleeing = False
 
@@ -401,4 +402,21 @@ class HerdingEnv(gym.Env):
             pos   = np.clip(pos + (force / mag) * speed * self.DT,
                             -self.FIELD, self.FIELD)
 
+        # Pen solid wall collision — mirrors Webots geometry.
+        # The pen has THREE solid walls: west (x=PEN_X[0]), east (x=PEN_X[1]),
+        # south (y=PEN_Y[0]).  The NORTH face (y=PEN_Y[1]=-8) is the open entrance.
+        # Sheep may only enter through the north face; crossing a solid wall is blocked.
+        px0, px1 = self.PEN_X[0], self.PEN_X[1]
+        py0, py1 = self.PEN_Y[0], self.PEN_Y[1]
+        entered_from_north = (
+            old_pos[1] >= py1 and pos[1] < py1 and px0 < pos[0] < px1
+        )
+        if not entered_from_north:
+            # Block crossing through west wall from outside
+            if old_pos[0] < px0 <= pos[0] and py0 < pos[1] < py1:
+                pos = np.array([px0 - 1e-3, pos[1]], dtype=np.float32)
+            # Block crossing through east wall from outside
+            if old_pos[0] > px1 >= pos[0] and py0 < pos[1] < py1:
+                pos = np.array([px1 + 1e-3, pos[1]], dtype=np.float32)
+
         return pos.astype(np.float32)