BurnyCoder commited on
Commit
ea2cb7f
·
verified ·
1 Parent(s): 8a60395

generalist/seed1: trained

Browse files
Files changed (5) hide show
  1. README.md +47 -9
  2. final.pt +3 -0
  3. run.json +67 -2
  4. train_metrics.csv +0 -0
  5. val_metrics.csv +12 -0
README.md CHANGED
@@ -13,9 +13,9 @@ tags:
13
 
14
  # racingmania-generalist-seed1
15
 
16
- **Generalist (PPO)**, seed 1: a run of the RacingMania research project.
17
 
18
- Status: **planned** (registered; code and config pushed before training); physics version v1; card generated 2026-11-14 from the result files.
19
 
20
  RacingMania is an original 3D time-attack racing game for the browser (three.js + cannon-es `RaycastVehicle`) with reinforcement-learning drivers trained on a batched PyTorch port of the same car physics. Code, paper and full results: <https://github.com/BurnyCoder/reinforcement-learning-in-trackmania-like-racing-game>.
21
 
@@ -26,12 +26,12 @@ RacingMania is an original 3D time-attack racing game for the browser (three.js
26
  | Algorithm | PPO |
27
  | Architecture | MLP 2x256 tanh (actor and critic), two 3-way heads |
28
  | Objective | centreline progress |
29
- | Observation | car state + look-ahead points (default) |
30
- | Training data | generated training pool |
31
- | Start states | standing + random centreline |
32
- | Agent steps trained | - |
33
- | Wall-clock hours | - |
34
- | Git commit | - |
35
 
36
  ## Evaluation
37
 
@@ -73,7 +73,45 @@ To watch it drive, open the game (`web/index.html`, or the GitHub Pages site) an
73
 
74
  ## Training configuration (from run.json)
75
 
76
- _Not logged yet._
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77
 
78
  ## Licence and citation
79
 
 
13
 
14
  # racingmania-generalist-seed1
15
 
16
+ **Generalist (PPO)**, seed 1: Generalist PPO agent trained on the whole generated training pool
17
 
18
+ Status: **trained** (trained to its step budget, not yet evaluated); physics version v1; card generated 2026-11-14 from the result files.
19
 
20
  RacingMania is an original 3D time-attack racing game for the browser (three.js + cannon-es `RaycastVehicle`) with reinforcement-learning drivers trained on a batched PyTorch port of the same car physics. Code, paper and full results: <https://github.com/BurnyCoder/reinforcement-learning-in-trackmania-like-racing-game>.
21
 
 
26
  | Algorithm | PPO |
27
  | Architecture | MLP 2x256 tanh (actor and critic), two 3-way heads |
28
  | Objective | centreline progress |
29
+ | Observation | car state + 20 look-ahead points |
30
+ | Training data | 48,949 generated tracks |
31
+ | Start states | standing + random centreline (p=0.5) |
32
+ | Agent steps trained | 1,000,341,504 |
33
+ | Wall-clock hours | 1.92 |
34
+ | Git commit | `a861208dbcd896c0f7fe896657c5ee2ca2f45920` |
35
 
36
  ## Evaluation
37
 
 
73
 
74
  ## Training configuration (from run.json)
75
 
76
+ | Setting | Value |
77
+ |---|---|
78
+ | `total_steps` | `1000000000` |
79
+ | `num_envs` | `16384` |
80
+ | `num_steps` | `32` |
81
+ | `lr` | `0.001` |
82
+ | `adaptive_lr` | `True` |
83
+ | `desired_kl` | `0.01` |
84
+ | `lr_min` | `1e-05` |
85
+ | `lr_max` | `0.01` |
86
+ | `gamma` | `0.99` |
87
+ | `gae_lambda` | `0.95` |
88
+ | `num_minibatches` | `4` |
89
+ | `update_epochs` | `5` |
90
+ | `clip_coef` | `0.2` |
91
+ | `ent_coef` | `0.01` |
92
+ | `vf_coef` | `1.0` |
93
+ | `max_grad_norm` | `1.0` |
94
+ | `weight_decay` | `0.0` |
95
+ | `hidden` | `[256, 256]` |
96
+ | `action_mode` | `factored` |
97
+ | `eval_every_minutes` | `10.0` |
98
+ | `checkpoint_every_minutes` | `10.0` |
99
+ | `rotate_tracks_every` | `20` |
100
+ | `loaded_tracks` | `4096` |
101
+ | `seed` | `1` |
102
+ | `compile` | `True` |
103
+ | `env.random_start_prob` | `0.5` |
104
+ | `env.random_start_speed` | `[0.0, 30.0]` |
105
+ | `env.random_start_lateral` | `0.5` |
106
+ | `env.max_steps` | `1500` |
107
+ | `env.reward_scale` | `0.1` |
108
+ | `env.window_k` | `20` |
109
+ | `env.window_stride` | `3` |
110
+ | `env.blind` | `False` |
111
+ | `env.time_cost` | `0.0` |
112
+ | `env.fall_penalty` | `0.0` |
113
+ | `env.hint_weight` | `0.0` |
114
+ | `env.scripted_starts` | `False` |
115
 
116
  ## Licence and citation
117
 
final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:863cbbb64bf8db342ebd20fddb226f7851a2053e3ad6c6a9898b40c89d1b59e5
3
+ size 870476
run.json CHANGED
@@ -2,7 +2,72 @@
2
  "experiment": "generalist",
3
  "seed": 1,
4
  "track": null,
5
- "status": "planned",
6
  "phase": "2-single-seed",
7
- "updated": "2026-11-14T11:56:31.935449+00:00"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  }
 
2
  "experiment": "generalist",
3
  "seed": 1,
4
  "track": null,
5
+ "status": "trained",
6
  "phase": "2-single-seed",
7
+ "updated": "2026-11-14T13:52:46.891775+00:00",
8
+ "algo": "ppo",
9
+ "description": "Generalist PPO agent trained on the whole generated training pool",
10
+ "git_commit": "a861208dbcd896c0f7fe896657c5ee2ca2f45920",
11
+ "physics": "v1",
12
+ "hardware": {
13
+ "platform": "Linux-6.18.58.1-microsoft-standard-WSL2-x86_64-with-glibc2.39",
14
+ "python": "3.13.18",
15
+ "torch": "2.15.0+cu132",
16
+ "gpu": "NVIDIA GeForce RTX 5070 Laptop GPU"
17
+ },
18
+ "ppo": {
19
+ "total_steps": 1000000000,
20
+ "num_envs": 16384,
21
+ "num_steps": 32,
22
+ "lr": 0.001,
23
+ "adaptive_lr": true,
24
+ "desired_kl": 0.01,
25
+ "lr_min": 1e-05,
26
+ "lr_max": 0.01,
27
+ "gamma": 0.99,
28
+ "gae_lambda": 0.95,
29
+ "num_minibatches": 4,
30
+ "update_epochs": 5,
31
+ "clip_coef": 0.2,
32
+ "ent_coef": 0.01,
33
+ "vf_coef": 1.0,
34
+ "max_grad_norm": 1.0,
35
+ "weight_decay": 0.0,
36
+ "hidden": [
37
+ 256,
38
+ 256
39
+ ],
40
+ "action_mode": "factored",
41
+ "eval_every_minutes": 10.0,
42
+ "checkpoint_every_minutes": 10.0,
43
+ "rotate_tracks_every": 20,
44
+ "loaded_tracks": 4096,
45
+ "seed": 1,
46
+ "compile": true
47
+ },
48
+ "env": {
49
+ "random_start_prob": 0.5,
50
+ "random_start_speed": [
51
+ 0.0,
52
+ 30.0
53
+ ],
54
+ "random_start_lateral": 0.5,
55
+ "max_steps": 1500,
56
+ "reward_scale": 0.1,
57
+ "window_k": 20,
58
+ "window_stride": 3,
59
+ "blind": false,
60
+ "time_cost": 0.0,
61
+ "fall_penalty": 0.0,
62
+ "hint_weight": 0.0,
63
+ "scripted_starts": false
64
+ },
65
+ "n_train_tracks": 48949,
66
+ "vehicle_overrides": {},
67
+ "checkpoint_dir": "/mnt/d/AI/racingmania/checkpoints/generalist/seed1",
68
+ "tensorboard_dir": "/mnt/d/AI/racingmania/tensorboard/generalist/seed1",
69
+ "global_step": 1000341504,
70
+ "iterations": 1908,
71
+ "wall_hours": 1.9208961590131124,
72
+ "last_session_hours": 1.921039325926039
73
  }
train_metrics.csv ADDED
The diff for this file is too large to render. See raw diff
 
val_metrics.csv ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ sws,finish_rate,speed_ratio_finished,n_tracks,flags_overspeed,flags_energy_gain,iteration,global_step,wall_hours
2
+ 0.0,0.0,nan,197,0,0,0,0,0.0
3
+ 1.0633366484118099,0.7055837563451777,1.5070310772455147,197,0,0,190,99614720,0.16784764839543237
4
+ 1.222279800802202,0.8071065989847716,1.5143969858995834,197,0,0,447,234356736,0.34694840570290886
5
+ 1.235236382031222,0.8223350253807107,1.5021084398774736,197,0,0,649,340262912,0.5273539512024985
6
+ 1.2149324516693405,0.8071065989847716,1.5052936665337109,197,0,0,845,443023360,0.704637470708953
7
+ 1.2137514569703414,0.8071065989847716,1.503830421529291,197,0,0,1004,526385152,0.8858941390779284
8
+ 1.2266359276614214,0.817258883248731,1.5009147686291928,197,0,1,1163,609746944,1.0669237690501743
9
+ 1.2469413931465156,0.8324873096446701,1.497850332011363,197,0,0,1306,684720128,1.2468739506271151
10
+ 1.2210377541718485,0.817258883248731,1.4940648296388455,197,0,0,1450,760217600,1.4269774105813768
11
+ 1.2547547272829231,0.8578680203045685,1.4626430844658924,197,0,0,1594,835715072,1.6063099728027979
12
+ 1.221571630943686,0.8121827411167513,1.5040600705994134,197,0,0,1798,942669824,1.785360698832406