2de3085f13
chore(training): curriculum progress after curric-s5-aggression
2026-07-22 20:33:07 +01:00
fda784d50d
chore(training): Add curric-s5-aggression checkpoints, logs, and exported policy
2026-07-22 20:31:54 +01:00
fca6a46200
fix(training): correct stage-3 eval (locomotion-mask bugfix) and add grounded aggression stage
...
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.
2026-07-22 12:48:45 +01:00
cf4859e61c
chore(training): Add curric-s4-mechanics checkpoints, logs, and exported policy
2026-07-22 11:53:29 +01:00
8c15c466ef
fix(*): apply the locomotion mask during in-game/eval inference, not just training
...
AIShipController (eval + real gameplay) ran the raw policy output unmasked
regardless of allow_vertical/allow_pitch_roll, while ShipAIController
(training) correctly discarded those axes for grounded curriculum stages.
A grounded-trained model's untrained vertical/pitch-roll output reached the
ship as noise during eval, understating it against models that were never
handicapped this way.
2026-07-21 22:23:09 +01:00
dd2b3c570b
chore(training): curriculum progress after curric-s3-no_draws
2026-07-21 21:46:34 +01:00
bc644e0305
chore(training): Add curric-s3-no_draws checkpoints, logs, and exported policy
2026-07-21 21:44:51 +01:00
44b1384a97
chore(training): curriculum progress after curric-s2-defend
2026-07-21 19:19:44 +01:00
ffa9017cbd
chore(training): Add curric-s2-defend checkpoints, logs, and exported policy
2026-07-21 19:17:32 +01:00
7e67998a23
chore(training): curriculum progress after curric-s1-score
2026-07-21 16:52:20 +01:00
523a26aa60
chore(training): Add curric-s1-score checkpoints, logs, and exported policy
2026-07-21 16:50:08 +01:00
8e3fafcc8b
feat(*): add staged curriculum training with an automated stage-by-stage orchestrator
2026-07-21 12:38:51 +01:00
1d539cc8c7
chore(training): Add run11 checkpoints, logs, and exported policy
2026-07-21 09:40:13 +01:00
fbf51eebe5
chore(training): Add run10 checkpoints, logs, and exported policy
2026-07-20 22:34:32 +01:00
7d69ac4a01
feat(*): retune scoring incentives and add finishing reps
2026-07-20 20:06:11 +01:00
e23d780402
chore(training): Add run09 checkpoints, logs, and exported policy
2026-07-20 19:09:32 +01:00
6cb5902eb6
feat(*): Add long training run script
2026-07-20 07:18:13 +01:00
48369c50dc
chore(training): Add run07 checkpoints, logs, and exported policy
2026-07-20 05:56:00 +01:00
b8f5f4938d
chore(training): Add run06 checkpoints, logs, and exported policy
2026-07-20 00:42:52 +01:00
dfb7285704
chore(training): Add run05 checkpoints, logs, and exported policy
2026-07-19 21:58:08 +01:00
d917d2742b
feat(training): Add next_run.sh to idempotently kick off the next run resuming from the newest final.zip
2026-07-19 19:16:01 +01:00
cd94252f92
chore(*): Remove run05 artifacts trained with collapsed exploration
2026-07-19 19:16:01 +01:00
c517a85508
chore(training): Add run05 checkpoints, logs, and exported policy
2026-07-19 19:11:31 +01:00
80e9b676d5
chore(training): Add run04 checkpoints, logs, and exported policy
2026-07-19 17:54:36 +01:00
4e2d406aa2
chore(*): Remove run03 artifacts trained against the floor-taxed reward
2026-07-19 15:35:55 +01:00
5973d0893b
chore(training): Add run03 checkpoints, logs, and exported policy
2026-07-19 15:21:06 +01:00
772f98b7fe
feat(*): Fix exported-policy action order to gymnasium's sorted-key layout, add wall-contact penalty and stronger ball-touch reward, and wire Spectate to run01 vs run02
2026-07-19 13:21:14 +01:00
6f1a840249
chore(training): Add run02 checkpoints, logs, and exported policy
2026-07-19 12:41:44 +01:00
5480b3cf05
feat(*): Add --n-steps and --batch-size flags to train.py, applied on resume as well
2026-07-19 10:41:11 +01:00
6406867911
chore(*): Remove smoke-run and aborted run03 artifacts
2026-07-19 10:36:51 +01:00
f4722dbd8d
feat(*): Add idempotent tmux-based training scripts for the Linux box
2026-07-19 10:27:21 +01:00
1b3c2e063d
chore(training): Track training artifacts in git, add idempotent Linux setup/run scripts, and commit run01 results
2026-07-19 10:22:22 +01:00
07217c3517
feat(*): Add bot-vs-bot Spectate mode with main-menu entry, entropy-control flags (--ent-coef, --reset-std) for resumed training runs, and a Linux/3090 remote-training guide (TRAINING_LINUX.md)
2026-07-19 10:10:32 +01:00
85f96eb15e
feat(*): Add self-play RL training pipeline with PPO trainer, in-game GDScript policy inference, and bot opponent support in Match mode
2026-07-18 19:32:51 +01:00
CosmicClash Training Bot