Skip to content

Commit 343a769

Browse files
salim4nkhayonclaude
authored
docs(space-dock): mention de maquette provisoire, virgule EN, « potentiel » non prouvé retiré (#99)
* docs(space-dock): restore technical-mockup notice, fix EN comma, drop unproven "potential" claim Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * ci: relancer la CI (ppo.test.ts convergence instable, sans lien avec ce correctif de texte) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: khayon <khayon@ignitionai.local> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
1 parent 26344aa commit 343a769

4 files changed

Lines changed: 4 additions & 4 deletions

File tree

‎packages/web/blog/space-dock-physique-pd.en.mdx‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -145,7 +145,7 @@ In the demo, `TrainingSession` runs these transitions in 12 ms chunks so the sce
145145

146146
The curve shows the mean reward of the last ten episodes played in your browser; the docking rate counts episodes that actually succeeded.
147147

148-
What we measured outside the browser (Node.js, no rendering, seed 118, 20 greedy episodes from the standard start every 2,000 steps): without the curriculum and the potential, no docking in 40,000 steps. With the demo's recipe, the network docks 20 episodes out of 20 at 32,000 steps, then learning degrades (at most 1 in 20 between 34,000 and 40,000 steps). One run, one seed: it is a result, not a guarantee. In the browser, the demo was measured at between 65 and 104 steps per second in our acceptance test (Chromium); this varies by machine. At that rate, 32,000 steps take at least five minutes (32,000 / 104 ≈ 5 min).
148+
What we measured outside the browser (Node.js, no rendering, seed 118, 20 greedy episodes from the standard start every 2,000 steps): without the curriculum, no docking in 40,000 steps. With the demo's recipe, the network docks 20 episodes out of 20 at 32,000 steps, then learning degrades (at most 1 in 20 between 34,000 and 40,000 steps). One run, one seed: it is a result, not a guarantee. In the browser, the demo was measured at between 65 and 104 steps per second in our acceptance test (Chromium); this varies by machine. At that rate, 32,000 steps take at least five minutes (32,000 / 104 ≈ 5 min).
149149

150150
Hence the "Charger un agent pré-entraîné" (load a pre-trained agent) button. We trained that network offline with the same recipe: 32,000 steps, seed 118. Before publishing it, we evaluated it with no exploration on 50 unseen episodes from the standard start: 50 dockings, mean final distance 0.29 m. Live training remains available; "Réinitialiser le réseau" starts again from initial weights.
151151

‎packages/web/blog/space-dock-physique-pd.fr.mdx‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -145,7 +145,7 @@ Dans la démo, `TrainingSession` exécute ces transitions par tranches de 12 ms
145145

146146
La courbe affiche la récompense moyenne des dix derniers épisodes joués dans votre navigateur ; le taux d'arrimage compte les épisodes réellement réussis.
147147

148-
Ce que nous avons mesuré, hors navigateur (Node.js, sans rendu, graine 118, 20 épisodes greedy depuis le départ standard tous les 2 000 pas) : sans curriculum ni potentiel, aucun arrimage en 40 000 pas. Avec la recette de la démo, le réseau arrime 20 épisodes sur 20 à 32 000 pas, puis l'apprentissage se dégrade (au plus 1 sur 20 entre 34 000 et 40 000 pas). Un seul entraînement, une seule graine : c'est un résultat, pas une garantie. Dans le navigateur, la démo a été mesurée entre 65 et 104 pas par seconde dans notre recette (Chromium) ; cela varie selon la machine. À ce rythme, 32 000 pas demandent au moins cinq minutes (32 000 / 104 ≈ 5 min).
148+
Ce que nous avons mesuré, hors navigateur (Node.js, sans rendu, graine 118, 20 épisodes greedy depuis le départ standard tous les 2 000 pas) : sans curriculum, aucun arrimage en 40 000 pas. Avec la recette de la démo, le réseau arrime 20 épisodes sur 20 à 32 000 pas, puis l'apprentissage se dégrade (au plus 1 sur 20 entre 34 000 et 40 000 pas). Un seul entraînement, une seule graine : c'est un résultat, pas une garantie. Dans le navigateur, la démo a été mesurée entre 65 et 104 pas par seconde dans notre recette (Chromium) ; cela varie selon la machine. À ce rythme, 32 000 pas demandent au moins cinq minutes (32 000 / 104 ≈ 5 min).
149149

150150
D'où le bouton « Charger un agent pré-entraîné ». Ce réseau a été entraîné hors ligne par nos soins avec la même recette : 32 000 pas, graine 118. Avant publication, nous l'avons évalué sans exploration sur 50 épisodes inédits depuis le départ standard : 50 arrimages, distance finale moyenne 0,29 m. L'entraînement en direct reste disponible ; « Réinitialiser le réseau » repart de poids initiaux.
151151

‎packages/web/data/blog.en.json‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22
{
33
"slug": "space-dock-physique-pd",
44
"title": "Space Dock: docking cargo with Rapier, a reference PD controller and a trainable DQN",
5-
"description": "A tug and its cargo physically linked: manual controls, an untrained PD controller, docking criteria in-browser DQN training and an offline pre-trained agent, with its measurements.",
5+
"description": "A tug and its cargo physically linked: manual controls, an untrained PD controller, docking criteria, in-browser DQN training and an offline pre-trained agent, with its measurements.",
66
"imageAlt": "The tug and its cargo facing the docking ring in the industrial orbital hangar."
77
},
88
{

‎packages/web/data/demos.json‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -114,7 +114,7 @@
114114
{
115115
"slug": "space-dock",
116116
"title": "Space Dock — technical preview",
117-
"description": "Tow a rigidly linked cargo with real Rapier forces and collisions. Train a DQN in your browser or load an agent we pre-trained offline, watch it fly the tug without exploration, and compare with manual controls and an untrained scripted PD reference.",
117+
"description": "Temporary technical mockup: tow a rigidly linked cargo with real Rapier forces and collisions. Train a DQN in your browser or load an agent we pre-trained offline, watch it fly the tug without exploration, and compare with manual controls and an untrained scripted PD reference; final visuals pending.",
118118
"tech": "R3F · Rapier",
119119
"algos": "DQN · Manuel · PD non entraîné",
120120
"accent": "#7ac5f0",

0 commit comments

Comments
 (0)