chore(training): add air_touch_bonus_weight and restart stage-5 intercepts

air_approach_weight alone didn't move productive_air_touch_fraction after a
further 180M steps (360M cumulative across all six Stage-5 attempts): an
unredirected air-intercept ball falls short of the goal from gravity and
just lands on the floor, so the already-solved ground game collects the
same episode reward whether or not anything touched the ball in the air.
air_touch_bonus_weight adds a conjunctive event bonus on top of
ball_touch_reward for a touch that's both genuinely aerial and
goal-directed, targeting the actual measured behaviour instead of only the
approach to it.

Game/scripts/ship_ai_controller.gd

@@ -60,6 +60,29 @@ extends AIController3D
 # vector rather than the planar-only one, since a real aerial requires
 # pitching away from level.
 @export var air_approach_weight := 0.0
+# Event bonus, conjunctive with the same goal-direction alignment already
+# gating ball_touch_reward: extra payout for a touch that is BOTH genuinely
+# aerial (ball.y > AIR_TOUCH_HEIGHT, matching the productive_air_touch_
+# fraction telemetry definition exactly) AND goal-directed. air_approach_
+# weight alone did not move productive_air_touch_fraction (still 0.0 after
+# a further 180M steps, 360M cumulative) because nothing in the reward ever
+# made touching the ball while still airborne worth more than the
+# alternative every policy already had available for free: let gravity pull
+# an unredirected air-intercept ball back down (it falls well short of the
+# goal's ~0-1.5m height band over the required flight distance, so it does
+# not auto-score) and then collect the same goal_reward/ball_touch_reward
+# via the already-dominant, already-solved ground game once it lands. This
+# is deliberately NOT the standalone height-only bonus generation 4 ruled
+# out (see TRAINING.md's "why no air-touch reward" note, added to close the
+# RLGym wall-bounce exploit): it only pays scaled by the same alignment
+# dot-product as the base term, so batting the ball in a non-productive
+# direction earns nothing extra, same anti-farming shape as ball_touch_
+# reward itself. Also safe from that specific exploit on distributional
+# grounds: air-intercept spawns are central (x in [-8,8], z in [-10,10],
+# arena half-extents 18/27) and air-drill spawns keep AIR_DRILL_BALL_WALL_
+# CLEARANCE from every wall, so neither state can be solved by bouncing off
+# one.
+@export var air_touch_bonus_weight := 0.0
 @export var ball_velocity_to_goal_weight := 0.004
 # Per-tick penalty scaled by distance to the ball (full value at the arena's
 # far diagonal, 0 on top of the ball). Run04 lesson: with idling worth a flat
@@ -504,14 +527,16 @@ func _on_ship_body_entered(body: Node) -> void:
 	var to_goal := attack_goal_position - ball.global_position
 	if to_goal.length_squared() > 0.0001 and ball.linear_velocity.length_squared() > 0.0001:
 		alignment = clampf(ball.linear_velocity.normalized().dot(to_goal.normalized()), 0.0, 1.0)
-	reward += ball_touch_reward * lerpf(ball_touch_direction_floor, 1.0, alignment)
+	var touch_payout := ball_touch_reward * lerpf(ball_touch_direction_floor, 1.0, alignment)
+	if air_touch_bonus_weight > 0.0 and ball.global_position.y > AIR_TOUCH_HEIGHT:
+		touch_payout += air_touch_bonus_weight * alignment
+	reward += touch_payout
 	_ticks_since_ball_touch = 0
 
-	# Telemetry only (see get_info's air_touch_fraction) — not a reward term.
+	# air_touch_fraction/productive_air_touch_fraction (see get_info) share
-	# Generation 4 deliberately doesn't reward high touches directly (see
+	# their AIR_TOUCH_HEIGHT/alignment definitions 1:1 with air_touch_bonus_
-	# TRAINING.md's "why no air-touch reward" note); this just measures
+	# weight above by design — the reward now targets exactly the behaviour
-	# whether the air-drill state setter is producing genuine aerial
+	# the telemetry measures.
-	# contests, so a future decision to add one is data-driven.
 	_touches += 1
 	if ball.global_position.y > AIR_TOUCH_HEIGHT:
 		_air_touches += 1

Game/scripts/training_mode.gd

@@ -265,7 +265,7 @@ const TRAINING_MODE_OVERRIDES := [
 const SHIP_AI_OVERRIDES := [
 	"ball_touch_reward", "ball_touch_cooldown_ticks", "ball_touch_direction_floor",
 	"velocity_to_ball_weight", "ball_velocity_to_goal_weight", "ball_distance_penalty",
-	"forward_velocity_to_ball_weight", "air_approach_weight", "wall_contact_penalty", "tilt_penalty",
+	"forward_velocity_to_ball_weight", "air_approach_weight", "air_touch_bonus_weight", "wall_contact_penalty", "tilt_penalty",
 	"ground_tilt_penalty", "non_forward_penalty", "grounded_upright_reward",
 	"speed_reward_weight", "time_penalty", "airborne_penalty",
 ]
@@ -324,6 +324,7 @@ func _ai_default(name: String) -> Variant:
 		"velocity_to_ball_weight": return 0.02
 		"forward_velocity_to_ball_weight": return 0.0
 		"air_approach_weight": return 0.0
+		"air_touch_bonus_weight": return 0.0
 		"ball_velocity_to_goal_weight": return 0.004
 		"ball_distance_penalty": return 0.002
 		"wall_contact_penalty": return 0.0025

TRAINING.md

@@ -662,6 +662,34 @@ were deleted and Stage 5 restarts from Stage 4's checkpoint with the new term,
 same reasoning as every previous mechanism change: don't resume a policy
 shaped by an absent term into one where it now exists.
 
+**`air_approach_weight` alone did not fix it.** A further three attempts
+(180M more steps, 360M cumulative across all six Stage-5 attempts, 2026-08-18
+to 2026-08-19) closed with `productive_air_touch_fraction` still exactly 0.0
+and `air_touch_fraction` still at noise level, while `goal_rate` kept clearing
+its (lower) floor. Working out the actual physics instead of retuning another
+number found the real gap: an unredirected `_place_air_intercept` ball
+(spawned 6-12m up, aimed at a goal whose collision box sits at ~0-1.5m) sags
+well short of the goal from gravity alone over the required flight
+distance — it does not auto-score — so it just falls to the floor, where the
+already-solved ground game collects the exact same `goal_reward`/
+`ball_touch_reward` regardless of whether anything touched it in the air.
+Nothing in the reward ever made a genuinely aerial touch worth more than
+waiting the second or two for the ball to land, so `air_approach_weight`'s
+dense closing-speed shaping had nothing to reinforce — a second,
+independent missing-incentive gap in the same stage, not a training-duration
+problem. `air_touch_bonus_weight` (`ship_ai_controller.gd`) closes it
+directly: an event bonus on top of `ball_touch_reward`, paid only for a
+touch that is both above `AIR_TOUCH_HEIGHT` and goal-directed, scaled by the
+exact same alignment factor already gating the base touch reward —
+conjunctive, not standalone, so it targets exactly the behaviour
+`productive_air_touch_fraction` measures without reopening the RLGym
+wall-bounce exploit the original "no standalone air-touch reward" decision
+(above) was written to avoid: air-intercept/air-drill spawns are kept away
+from every wall by construction. Set to 0.5 (roughly `ball_touch_reward`'s
+own magnitude) and folded into `HANDLING_REWARD_FLAGS`. The three blocked
+attempts were deleted and Stage 5 restarts from Stage 4's checkpoint again
+with both terms active.
+
 Stage 6's `league` opponent mode samples a historical exported policy at each
 episode reset. Each later stage preserves the preceding shaping and adds one
 new difficulty.

training/eval_history.json

@@ -873,5 +873,32 @@
       "team_1": 54
     },
     "win_rate_a": 0.58
+  },
+  {
+    "timestamp": "2026-08-19T21:34:06+00:00",
+    "model_a": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260819-1321-gen5-s5-intercepts-retry2.json",
+    "model_b": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260816-2126-gen5-s4-handling-retry2.json",
+    "seed": 1,
+    "episodes": 100,
+    "wins_a": 56,
+    "wins_b": 31,
+    "draws": 13,
+    "side_results": {
+      "a_team_0": {
+        "wins_a": 25,
+        "wins_b": 16,
+        "draws": 9
+      },
+      "a_team_1": {
+        "wins_a": 31,
+        "wins_b": 15,
+        "draws": 4
+      }
+    },
+    "physical_team_wins": {
+      "team_0": 40,
+      "team_1": 47
+    },
+    "win_rate_a": 0.56
   }
 ]

training/generation5.py

@@ -181,10 +181,40 @@ STANDING_ARGS = ["--ent-coef", "0.01", "--entropy-floor"]
 # than continuing retry2's, same reasoning as every previous mechanism
 # change in this file: don't resume a policy shaped by an absent term into
 # one where it now exists.
+#
+# Round 8 (2026-08-19): air_approach_weight alone did not move the needle
+# either -- another full 180M-step chain (3 more attempts, 360M cumulative
+# across all six Stage-5 attempts) closed with productive_air_touch_fraction
+# still exactly 0.0 and air_touch_fraction at noise level, while goal_rate
+# kept passing its (lower) floor. Working out the physics instead of just
+# re-tuning a number found why: an unredirected air-intercept ball (spawned
+# 6-12m up, aimed at a goal whose collision box sits at ~0-1.5m) sags well
+# short of the goal from gravity alone over the required flight distance --
+# it does not auto-score -- so it simply falls to the floor, and the
+# already-solved ground game (forward_velocity_to_ball_weight, ball_touch_
+# reward, goal_reward) collects the exact same total episode reward either
+# way. Nothing ever made touching the ball while it was still genuinely
+# airborne worth more than waiting the second or two for it to land, so
+# air_approach_weight's dense closing-speed shaping had nothing to reinforce
+# -- nowhere near a training-duration problem, a second missing-incentive
+# gap in the same stage.
+#
+# air_touch_bonus_weight (ship_ai_controller.gd) closes it directly: an
+# event bonus on top of ball_touch_reward, paid only for a touch that is
+# both above AIR_TOUCH_HEIGHT and goal-directed, scaled by the exact same
+# alignment factor already gating the base touch reward -- conjunctive, not
+# standalone, so it can't be farmed by batting the ball in a useless
+# direction, and it targets exactly the behaviour productive_air_touch_
+# fraction measures instead of only the approach to it. Set to 0.5 (roughly
+# ball_touch_reward's own magnitude, so a fully-aligned aerial touch pays
+# ~1.7x a fully-aligned ground one). Also folded into HANDLING_REWARD_FLAGS
+# so Stage 6 inherits it. Restarts Stage 5 from Stage 4's checkpoint again,
+# same reasoning as every prior mechanism change here.
 HANDLING_REWARD_FLAGS = [
     "--velocity-to-ball-weight", "0.04",
     "--forward-velocity-to-ball-weight", "0.15",
     "--air-approach-weight", "0.15",
+    "--air-touch-bonus-weight", "0.5",
     "--ball-distance-penalty", "0.01",
     "--ball-touch-reward", "0.7",
     "--ball-velocity-to-goal-weight", "0.06",

training/generation5_state.json

@@ -1,6 +1,6 @@
 {
   "stage_index": 1,
-  "attempt": 2,
+  "attempt": 0,
   "status": "in_progress",
   "log": [
     {
@@ -258,171 +258,6 @@
         ],
         "reason": "Human override. Only miss was the training goal-rate floor, at 0.7731 vs 0.80; every evaluation gate passed (65-22-13 vs promoted/easy.json, 87% non-draw, 12.6% physical-side imbalance) and the goal rate improved monotonically across all three attempts (0.537 -> 0.683 -> 0.773). Also promoted to Game/bots/promoted/medium.json on the same evidence. Stage 5 resumes from this checkpoint."
       }
-    },
-    {
-      "stage_index": 1,
-      "stage_number": 5,
-      "stage_name": "intercepts",
-      "experiment": "20260818-1903-gen5-s5-intercepts",
-      "attempt": 0,
-      "telemetry_tail": {
-        "rollout/air_touch_fraction": 0.00011999999731779099,
-        "rollout/airborne_fraction": 0.2547863810658455,
-        "rollout/ep_len_mean": 97.30996008300781,
-        "rollout/ep_rew_mean": 16.281804374694826,
-        "rollout/forward_motion_fraction": 0.33076866614818573,
-        "rollout/goal_rate": 0.8198399951457978,
-        "rollout/grounded_upright_fraction": 0.1814400002658367,
-        "rollout/mean_altitude": 3.249552144050598,
-        "rollout/productive_air_touch_fraction": 0.0,
-        "rollout/upright_fraction": 0.7921435704231262,
-        "rollout/vertical_thrust_mean": 0.06410999808754772
-      },
-      "telemetry_failures": [
-        "rollout/productive_air_touch_fraction=0.0000 < 0.0050"
-      ],
-      "evaluation_goal_failures": [],
-      "side_balance_failures": [],
-      "eval": {
-        "timestamp": "2026-08-19T03:12:42+00:00",
-        "model_a": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260818-1903-gen5-s5-intercepts.json",
-        "model_b": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260816-2126-gen5-s4-handling-retry2.json",
-        "seed": 1,
-        "episodes": 100,
-        "wins_a": 55,
-        "wins_b": 30,
-        "draws": 15,
-        "side_results": {
-          "a_team_0": {
-            "wins_a": 25,
-            "wins_b": 17,
-            "draws": 8
-          },
-          "a_team_1": {
-            "wins_a": 30,
-            "wins_b": 13,
-            "draws": 7
-          }
-        },
-        "physical_team_wins": {
-          "team_0": 38,
-          "team_1": 47
-        },
-        "win_rate_a": 0.55
-      },
-      "evals": [
-        {
-          "timestamp": "2026-08-19T03:12:42+00:00",
-          "model_a": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260818-1903-gen5-s5-intercepts.json",
-          "model_b": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260816-2126-gen5-s4-handling-retry2.json",
-          "seed": 1,
-          "episodes": 100,
-          "wins_a": 55,
-          "wins_b": 30,
-          "draws": 15,
-          "side_results": {
-            "a_team_0": {
-              "wins_a": 25,
-              "wins_b": 17,
-              "draws": 8
-            },
-            "a_team_1": {
-              "wins_a": 30,
-              "wins_b": 13,
-              "draws": 7
-            }
-          },
-          "physical_team_wins": {
-            "team_0": 38,
-            "team_1": 47
-          },
-          "win_rate_a": 0.55
-        }
-      ],
-      "decision": "fail"
-    },
-    {
-      "stage_index": 1,
-      "stage_number": 5,
-      "stage_name": "intercepts",
-      "experiment": "20260819-0412-gen5-s5-intercepts-retry1",
-      "attempt": 1,
-      "telemetry_tail": {
-        "rollout/air_touch_fraction": 0.00011999999731779099,
-        "rollout/airborne_fraction": 0.25744442877173424,
-        "rollout/ep_len_mean": 112.18267991638183,
-        "rollout/ep_rew_mean": 20.06156894302368,
-        "rollout/forward_motion_fraction": 0.32834857150912283,
-        "rollout/goal_rate": 0.7362400003671646,
-        "rollout/grounded_upright_fraction": 0.17714000023156404,
-        "rollout/mean_altitude": 3.2925854358673097,
-        "rollout/productive_air_touch_fraction": 0.0,
-        "rollout/upright_fraction": 0.7877685700654984,
-        "rollout/vertical_thrust_mean": 0.09532099807504565
-      },
-      "telemetry_failures": [
-        "rollout/goal_rate=0.7362 < 0.7500",
-        "rollout/productive_air_touch_fraction=0.0000 < 0.0050"
-      ],
-      "evaluation_goal_failures": [],
-      "side_balance_failures": [],
-      "eval": {
-        "timestamp": "2026-08-19T12:21:42+00:00",
-        "model_a": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260819-0412-gen5-s5-intercepts-retry1.json",
-        "model_b": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260816-2126-gen5-s4-handling-retry2.json",
-        "seed": 1,
-        "episodes": 100,
-        "wins_a": 58,
-        "wins_b": 33,
-        "draws": 9,
-        "side_results": {
-          "a_team_0": {
-            "wins_a": 26,
-            "wins_b": 22,
-            "draws": 2
-          },
-          "a_team_1": {
-            "wins_a": 32,
-            "wins_b": 11,
-            "draws": 7
-          }
-        },
-        "physical_team_wins": {
-          "team_0": 37,
-          "team_1": 54
-        },
-        "win_rate_a": 0.58
-      },
-      "evals": [
-        {
-          "timestamp": "2026-08-19T12:21:42+00:00",
-          "model_a": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260819-0412-gen5-s5-intercepts-retry1.json",
-          "model_b": "/home/jcreek/ai-training/CosmicClash/Game/bots/20260816-2126-gen5-s4-handling-retry2.json",
-          "seed": 1,
-          "episodes": 100,
-          "wins_a": 58,
-          "wins_b": 33,
-          "draws": 9,
-          "side_results": {
-            "a_team_0": {
-              "wins_a": 26,
-              "wins_b": 22,
-              "draws": 2
-            },
-            "a_team_1": {
-              "wins_a": 32,
-              "wins_b": 11,
-              "draws": 7
-            }
-          },
-          "physical_team_wins": {
-            "team_0": 37,
-            "team_1": 54
-          },
-          "win_rate_a": 0.58
-        }
-      ],
-      "decision": "fail"
     }
   ]
 }

training/train.py

@@ -360,6 +360,12 @@ def parse_args():
         help="Aerial mirror of forward-velocity-to-ball-weight: high-altitude dense reward for "
         "nose-led 3D closing speed toward the ball",
     )
+    curriculum.add_argument(
+        "--air-touch-bonus-weight", type=float, default=None,
+        help="Event bonus on top of ball-touch-reward for a touch that is both genuinely aerial "
+        "(ball above AIR_TOUCH_HEIGHT) and goal-directed, scaled by the same alignment factor "
+        "as the base touch reward",
+    )
     curriculum.add_argument(
         "--ball-distance-penalty", type=float, default=None,
         help="Overrides ShipAIController.ball_distance_penalty (dense per-tick cost scaled by distance to the ball)",
@@ -425,6 +431,7 @@ def _curriculum_kwargs(args) -> dict:
         "ai_velocity_to_ball_weight": args.velocity_to_ball_weight,
         "ai_forward_velocity_to_ball_weight": args.forward_velocity_to_ball_weight,
         "ai_air_approach_weight": args.air_approach_weight,
+        "ai_air_touch_bonus_weight": args.air_touch_bonus_weight,
         "ai_ball_distance_penalty": args.ball_distance_penalty,
         "ai_ball_touch_reward": args.ball_touch_reward,
         "ai_airborne_penalty": args.airborne_penalty,