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.
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.
Generation 2's single "unmask" stage (flip vertical/pitch-roll locomotion
from grounded-only to full 3D in one step) failed 3 independent 240M-step
attempts, landing at a stable 32% / 28% / 31% win rate vs curric-s5-aggression
each time -- not noise, and not fixable by more training time (attempts 2-3
each continued the same checkpoint lineage for another full 240M steps with
zero improvement). Every attempt shows train/std collapsing from ~0.30 to
~0.13-0.15 within the first ~10% of steps and never recovering: the policy
locks the newly-opened axes back down before ever meaningfully exploring
them.
Replaces the boolean allow_vertical/allow_pitch_roll mask on ShipAIController
with float vertical_ramp/pitch_roll_ramp multipliers (0.0-1.0), scaling axis
effect in set_action() instead of gating it outright -- the action space
never changes shape, so checkpoints stay resumable across ramp values. The
single unmask stage in curriculum.py becomes 4: three ungated warmup stages
(25%/50%/75% authority, airborne_penalty ramping in step) that train,
checkpoint, and always advance with no eval gate, then the measured stage at
full authority -- same reference, opponent mode, and 240M budget as the 3
failed attempts, for a direct comparison. Adds a "gated" flag/branch to
main()'s loop for the ungated stages.
This is generation 3 of the curriculum; generation 2's state is archived to
curriculum_state_gen2.json (mirroring the earlier gen1 -> gen2 archival) and
curriculum_state.json resets fresh, since its stage 0 no longer means what it
used to. See TRAINING.md's "Generation 3" section for the full postmortem,
stage table, and the open question about whether scaling action effect in
Godot (which PPO's own entropy/exploration math never sees) actually
addresses the collapse.
The vendored godot_rl sync bridge (Game/addons/godot_rl_agents/sync.gd,
_training_process) snapshots each agent's info dict once per tick and its
own inline comment already flags that reset-timing path as incomplete
("NEEDS REFACTOR"); at least one agent's terminal-step info can arrive
without "goal_scored" at all. Indexing it directly crashed a training run
(20260729-0607-curric-s1-unmask-retry2) within minutes of starting. Skip
episodes missing the key instead of crashing training over a
monitoring-only metric.
Stage 5 (aggression) passed (41-47 vs grounded curric-s2-defend, within
the lenient gate but not yet a clear win). Rather than keep the locomotion
mask on indefinitely, stage 6 reopens full 3D controls on top of the
aggression retune and pairs it with a new dense airborne_penalty (scaled
by height above the floor) so the policy learns to prefer staying grounded
through incentives instead of a hard mask — same regime shift that
regressed stage 3, but this time with a mitigation and ~12x the training
time (~240M timesteps / ~24h vs ~20M / ~2h) to actually re-converge
instead of stalling mid-shift.
airborne_penalty follows the existing SHIP_AI_OVERRIDES pattern: default
0 (off) on ship_ai_controller.gd, exposed via train.py's new
--airborne-penalty flag, added to training_mode.gd's allow-list. Also adds
a per-stage timesteps override in curriculum.py (STAGES[n]["timesteps"])
since this is the first stage to need a different budget than the rest.
Re-ran stage-3 (curric-s3-no_draws vs curric-s2-defend) and the missing
stage-4 gate now that the locomotion-mask inference bugfix is in. Both
reverse or contradict the pre-fix bookkeeping: curric-s2-defend (grounded)
beats curric-s3-no_draws 60-26 and curric-s4-mechanics 57-24 when fairly
evaluated, so lifting the locomotion mask in stage 3 was a real regression
in floor play, not the improvement the buggy eval reported.
Adds a stage-5 "aggression" curriculum entry that resumes from stage 2
directly (via new resume_from_experiment/reference_experiment stage-dict
overrides in curriculum.py) instead of compounding the regression through
stages 3-4, keeps the locomotion mask on, and retunes ball-pursuit reward
weights for much more aggressive floor play. Extends train.py with the
three new --velocity-to-ball-weight/--ball-distance-penalty/--ball-touch-reward
flags needed to forward that retune to Godot's existing SHIP_AI_OVERRIDES.
curriculum_state.json and TRAINING.md are corrected/annotated in place
rather than silently rewritten, so the regression stays visible in history.
Josh Creek