Approach refinement

training/herding_env.py

@@ -61,12 +61,11 @@ class HerdingEnv(gym.Env):
     W_COMPLETE  = 100.0  # all sheep penned
     W_STEP_COST = 0.02   # time penalty — strong enough to punish doing nothing
     W_COMPACT   = 0.0    # reward for flock-radius reduction (off by default)
-    W_WALL_TOUCH = 0.15  # per-sheep max penalty at wall surface. Linear ramp
-                         # within WALL_TOUCH_BUFFER gives the RL agent a gradient
-                         # signal to avoid pinning sheep against pen walls.
-                         # 0.15 ≈ 7.5× step_cost — strong enough to shape behavior
-                         # without overwhelming progress reward.
-    WALL_TOUCH_BUFFER = 0.8   # metres from wall where penalty starts ramping
+    W_WALL_TOUCH = 0.04  # per-sheep max penalty at wall surface. Linear ramp
+                         # within WALL_TOUCH_BUFFER nudges the agent to avoid
+                         # pinning sheep against pen walls. 0.04 ≈ 2× step_cost
+                         # — noticeable but never dominates progress reward.
+    WALL_TOUCH_BUFFER = 0.3   # metres from wall where penalty starts ramping
     ALIGN_SHAPE = "standoff"   # "standoff" (peaks at IDEAL) | "near" (peaks at 0)
     ALIGN_GATED = True   # gate alignment on action magnitude
     ENTRY_AWARE = False  # When True, targets PEN_ENTRY (entrance face) instead