Adversarial review of the previous stage-4 retune found two problems:
non_forward_speed used planar_speed - forward_component, which under-charges
diagonal motion relative to true lateral speed (e.g. ~29% penalty at 45
degrees off the nose instead of the correct ~71%); fixed to the Pythagorean
magnitude for forward-facing angles, full speed for backward-facing ones.
Also, ground_tilt_penalty and non_forward_penalty only ever cost reward near
the floor with nothing offsetting them above it, which could teach a policy
that's still bad at ground handling to just avoid the floor rather than get
better at it. Added grounded_upright_reward (ship_ai_controller.gd) plus a
new ShipObservations.is_floor_contact helper for genuine belly-on-floor
contact detection, so grounding well while upright is the locally profitable
choice, not just the least-punished one.
Stage 4's upright/forward-motion telemetry plateaued flat across all three
blocked attempts because ground_tilt_penalty (0.003) was too weak to matter
and nothing penalized sideways/reverse motion at all. Raise
ground_tilt_penalty to 0.05 and add a new non_forward_penalty term
(ship_ai_controller.gd) that directly costs non-forward planar velocity near
the floor, independent of the ball. Delete the three blocked attempts'
checkpoints/logs/exports and reset generation5_state.json so the next run
starts fresh from the Stage-3 foundation checkpoint instead of continuing
from the drifted retry2 weights.
assets/textures/planet_surface.png was byte-identical to and unused
in favor of assets/models/nebula_planet_planet_surface.png, the
sidecar Godot's glTF importer actually extracted from nebula_planet.glb
and uses at runtime. Deleted the orphan (+ its .import) and stopped
gen_planet_surface.py from writing it.
Also measured nebula_dust.gdshader's per-fragment depth-texture sample
cost (~0.07ms/frame at 500 particles, within noise) -- negligible, so
no budgeting concern for further particle work.
Records the drag/aperture/collider-bake/beyond-1v1 items as done, with a
short note on what actually shipped (scoped up to 5v5 mid-implementation)
and the explicit non-goals (no 2v2+ curriculum, no team-size UI) so it's
clear what's still open.
Extends ShipObservations beyond the old self+1-opponent layout to padded
teammate/opponent arrays (MAX_TEAMMATES=4, MAX_OPPONENTS=5, SIZE=83),
zero-filling slots past the real roster size the same way the old single-
opponent slot was zero-filled when absent.
Slot stability across ticks requires a persistent identity: Ship gains
spawn_index (set once by GameMode.spawn_ship, never reassigned — there's no
despawn path anywhere in this codebase, so a roster is fixed for the whole
episode/match). ai_ship_controller.gd's opponent discovery is rewritten from
"first non-self ship" to classify every other ship by team and sort by
spawn_index; training_mode.gd/ship_ai_controller.gd carry the equivalent
sorted lists through the training path so both agree on slot assignment for
the same roster.
training_mode.gd and match_mode.gd both gain a team_size export (default 1,
so every existing curriculum script and match keeps today's 1v1 behaviour
unchanged). This is plumbing only: no 2v2+ curriculum or reward design, and
no match-mode UI to pick team size, has been done yet. The two checkpoints
in Game/bots/promoted/ are fitted to the old 35-float layout and are not
migrated — expected to go stale until the next training run.
Retraining from scratch removes the constraint that blocked these: the
existing checkpoints in Game/bots/promoted/ no longer need byte-identical
geometry.
- Fix the goal aperture constants (were 1.85/1.65, now match
objects/goal.tscn's actual sensor exactly at 1.75/1.5) so a ball crossing
the visible edge can't fail to score.
- Cut a real navigable hole in each end wall's collision
(_build_end_wall_colliders), replacing the previous solid box — the ball
now genuinely enters the net instead of triggering the sensor a hair
before hitting a solid wall. Correct for both FLOOR and ELEVATED goal
modes via _goal_surround_bounds. A separate, slightly wider
GOAL_VISUAL_APERTURE_* pair keeps the collision hole exact while still
giving goal.gd's bezel/rim frame clearance to be seen against the hull cut.
- Resize the play volume 1.5x (INNER_HALF_X/Z/HEIGHT 12/18/12 -> 18/27/18) to
comfortably fit a 5v5 roster: updates every hand-authored literal in
arena_boundary.tscn/arena_base.tscn/the elevated arena variants that
doesn't derive from those constants, adds 5 spawn markers per team, and
rescales ship_observations.gd's normalization scales and arena_02's
cosmetic decoration/nebula-dust shader uniforms to match.
- Bake the ~170 runtime-generated CollisionShape3D nodes into the scene via
a new bake_colliders()/tools/bake_arena_boundary.gd instead of rebuilding
them on every load — a real load-time cost repeated in every parallel
headless training env. Colliders must be direct children of the
StaticBody3D to register at all, so bake_colliders() parents them onto
self and tags them with a group for idempotent re-baking, rather than
grouping them under an intermediate container node. _ready() self-heals:
it skips regenerating only when an existing bake's goal_mode metadata
matches the current one, so the FLOOR-mode bake in the shared scene is
never silently reused by an ELEVATED arena variant.
_place_ships_random/_place_air_drill sampled each ship's randomized episode-
start position independently, so a team_size > 1 roster could spawn
interpenetrating (ships are ~1x1x4). Both now resample (up to 20 attempts,
matching the existing corner/fillet rejection-sampling pattern) against
every ship already placed that reset, rejecting anything within
MIN_SHIP_SEPARATION (4.5m, matching the arena spawn-marker spacing) of one.
The goal mouths are now a real navigable hole in the end walls, sized to the
ball rather than the ship — a ship's 1x1 cross-section fits through it, and
there's nothing behind the net to stop it. The escape failsafe previously
only existed in TrainingMode (where a physics regression just wastes
training time); now that any ship can genuinely fly out through an open
goal, every mode needs it or a stray ship/ball falls into the void with no
way back short of quitting. Moved up to GameMode as the shared default
_physics_process, removing TrainingMode's now-duplicate copy.
drag_coefficient/angular_drag/the idle angular-drag multiplier were applied
once per physics tick with no delta scaling, correct only because
project.godot never pins physics/common/physics_ticks_per_second and
Godot's default happens to be 60. _tick_scaled(k, state.step) makes the
decay rate invariant to tick rate instead. Also promotes the previously
hardcoded 0.9 idle angular-drag literal to an export, matching its sibling.
VecMonitor's info_keywords does a bare info[key] lookup on a completed
episode's terminal info dict and raises KeyError -- crashing the whole
training run -- if a key is ever absent. get_info() was only including
airborne_fraction/mean_altitude/vertical_thrust_mean when _telemetry_ticks
was nonzero and air_touch_fraction when _touches was nonzero; an episode
with zero ball touches (common, especially early in training) crashed on
the very first rollout in a smoke-test run. All four now always default to
0.0 rather than being conditionally present.
Found via TRAINING.md's Generation 4 validation ladder (rung 2, a 60k-step
smoke run) -- confirmed fixed by rerunning the same smoke run clean, then
export/evaluate parity (rungs 2-3) against Game/bots/promoted/easy.json.
All checkpoints/logs/exported policies here are a continuous-Gaussian,
31-input action/observation shape that generation 4's MultiDiscrete
redesign is structurally incompatible with -- nothing to resume from (see
the prior commit and TRAINING.md's "Generation 4" section). Game/bots/promoted/
(easy.json, and the new reference-grounded.json copied from
curric-s5-aggression before this) is untouched -- both remain valid,
playable evaluation opponents forever via PolicyNetwork's format-versioned
JSON despite their own checkpoints/generation being gone.
- training/checkpoints/*, training/logs/* removed (~3.5GB of working tree,
all generation 1-3 experiment runs).
- Game/bots/*.json flat dump removed (superseded exports; main_menu.gd's
Spectate dropdown will just be empty until the first generation-4 export).
- curriculum_state.json -> curriculum_state_gen3.json, archived alongside
the existing _gen1/_gen2 logs (all three are referenced as postmortem
evidence in TRAINING.md/curriculum.py). A fresh curriculum_state.json
will be created on the next curriculum.py run (load_state() already
handles a missing file).
training/eval_history.json is deliberately NOT reset -- it's the one
continuous cross-generation progress record.
NOT YET PUSHED: this needs the remote Linux training box quiesced first
(kill any active tmux session, confirm it's synced to origin) so its own
run_training.sh doesn't race a still-running job's final commit against
this deletion.
Three curriculum generations (2026-07-21 through 2026-08-04) all tried
gating *when* the policy could use vertical thrust/pitch-roll on top of a
continuous Gaussian action space, and all three failed the same way: PPO's
action-distribution std collapsed within ~10% of steps and never recovered,
landing at a 15-32% win rate vs the grounded reference regardless of
mechanism (hard mask, then a gradual ramp). Generation 3's final attempt
just landed at 24% — the worst of the three.
Root cause, verified against this project's own physics: hovering this ship
requires *holding* thrust.y ~= 0.408 continuously (mass 5.0, vertical_thrust
120, gravity 9.8). A collapsed near-zero-mean Gaussian can brush that value
but never sustain it long enough to earn the reward gradient that would
move the mean — no amount of gating *when* the axis acts fixes a problem in
*how* the policy represents a decision on it. This also independently found
and fixes a real bug: godot_rl never marks an episode timeout as a
truncation, so PPO was bootstrapping V(s)=0 on every 30s draw in every
generation to date.
- Game/scripts/ship_action_codec.gd (new): single source of truth for a
per-axis MultiDiscrete action space (7 heads, nvec [5,5,5,5,5,5,2]) shared
by training and in-game inference, replacing the continuous Gaussian.
thrust_y's bins are deliberately asymmetric so a random policy drifts
through the volume instead of floor-pinning. Legacy continuous decode
(ai_ship_controller.gd's old logic) preserved verbatim so every
pre-generation-4 export (e.g. Game/bots/promoted/easy.json) keeps working
unchanged via an optional "action_space" JSON field.
- ship_observations.gd: append own contact state (SIZE 31 -> 35, append-only)
so the value function can see what wall_contact_penalty fires on.
- ship_ai_controller.gd: action space/decode via the codec; drop the
vertical_ramp/pitch_roll_ramp mechanism entirely; tilt_penalty default
lowered 4x (aerial approaches require pitching); flight telemetry
(airborne_fraction, mean_altitude, air_touch_fraction, vertical_thrust_mean)
and truncation-snapshot fields on get_info().
- training_mode.gd: new air_drill_chance state-setter branch (ball spawned
high, ships low, kept clear of walls) so aerial practice is forced by the
environment instead of relying on reward-driven exploration alone; snapshot
terminal observations before a timeout reset for the truncation fix.
- cosmic_env.py: remap ShipAIController's truncated/terminal_obs info into
SB3's TimeLimit.truncated/terminal_observation keys.
- train.py: --reset-logits (+ --reset-logits-heads) replaces the
now-meaningless --reset-std; new EntropyFloorCallback (a persistent
per-rollout ent_coef controller replacing the one-shot std-reset shock)
and per-head entropy logging; FlightTelemetryCallback; --air-drill-chance/
--tilt-penalty flags; optional AbortIfCallback kill-criterion.
- export_policy.py: writes the action_space block for MultiDiscrete models;
index-level parity check (argmax per head) instead of comparing floats.
- curriculum.py: full rewrite — 3 stages (bootstrap/selfplay/gauntlet), no
grounded stage, full action space live from step 1; deletes generation
1-3's checkpoint-lineage machinery (nothing to resume from); final report
evaluates against both promoted/easy.json and the new
promoted/reference-grounded.json (a copy of curric-s5-aggression, the
strongest grounded-era artifact, kept as a fixed yardstick).
- run_training.sh/.gitignore: commit only final.zip, not the ~2400
intermediate checkpoint files a single stage was writing (~500MB ->
~0.2MB per run); requirements.txt pinned (behaviour here now depends on
specific library internals, not just public APIs).
- test_action_space.py (new): offline rung-0 check catching a head-order
mismatch before it silently corrupts 24h of training.
Validated: GDScript compiles clean (Godot --headless --import + script
validation), free_play.tscn and training.tscn both boot headless without
errors, offline action-space assertions pass. Not yet run: the actual
smoke-training/A-B validation ladder steps in TRAINING.md's "Generation 4"
section, before committing to the full ~32h curriculum.
See TRAINING.md's "Generation 4" section for the full design writeup.
ship.gd's controllerless path and player_ship_controller.gd each
allocated a fresh ShipAction every physics tick; ai_ship_controller.gd
and rl_ship_controller.gd already avoid this via a persistent member.
Convert both to reuse a member instance, matching the existing
full-field-overwrite convention (rather than +=/-= off a fresh zero).
Also drop the completed items from TODO.md.
HUDController found its ship via get_first_node_in_group("ship"), a group
that has 2+ members once a match has an AI opponent — it only worked
because the player ship happened to spawn first. spawn_camera_rig now
wires the HUD's ship the same way it already wires the camera rig's
target.
match_mode.gd fired the signal every frame while only the once-a-second
value is displayed, forcing the HUD to re-format and re-shape the label
each frame. Gate emission on the whole-second value changing, matching
ship.gd's threshold-gated telemetry discipline.
The coroutine stalls mid-countdown while the tree is paused for the
results screen, but _end_match unpauses before the deferred scene
change actually tears things down — letting it resume for a frame and
re-emit kickoff_countdown / unfreeze bodies in the dying scene.
CosmicClash Training Bot