Prist\u00e1tie bezpilotn\u00fdch lietadiel (UAV) na pohybuj\u00face sa platformy \u2013 od autodopravn\u00edkov cez paluby lod\u00ed a\u017e po mobiln\u00e9 roboty \u2013 predstavuje komplexn\u00fa \u00falohu s priese\u010dn\u00edkom percepcie, predikcie a riadenia. U\u010denie zosil\u0148ovan\u00edm (Reinforcement Learning, RL) pon\u00faka s\u013eubn\u00fa alternat\u00edvu k explicitn\u00e9mu n\u00e1vrhu regul\u00e1torov, preto\u017ee dok\u00e1\u017ee optimalizova\u0165 rozhodovanie v prostred\u00ed s ne\u00faplnou inform\u00e1ciou, stochastikou vetra a neline\u00e1rnymi obmedzeniami aktu\u00e1torov. Cie\u013eom je spo\u013eahliv\u00e9 a energeticky efekt\u00edvne dosadnutie pri minimaliz\u00e1cii pre\u0165a\u017een\u00ed, \u010dasu man\u00e9vru a rizika odtrhnutia kontaktu po dotyku (touchdown & securing<\/em>).<\/p>\n \u00dalohu modelujeme ako (\u010diasto\u010dne) pozorovate\u013en\u00fd Markovov rozhodovac\u00ed proces (POMDP) s komponentmi (S, A, T, R, O, \\u03B3)<\/em>:<\/p>\n Prirodzen\u00e9 je definova\u0165 stav v relat\u00edvnom r\u00e1mci platformy: $x_r, y_r, z_r, v_r, \\dot{\\psi}_r$, \u0161ikm\u00fd vietor a odhad trenia pri kontakte. Ak\u010dn\u00fd priestor je vhodn\u00e9 ohrani\u010di\u0165 fyzikou: $T \\in [T_{\\min}, T_{\\max}]$, r\u00fdchlosti n\u00e1klonov s limitmi slew-rate<\/em>. Bezpe\u010dnostn\u00e9 obmedzenia implementujeme ako tvrd\u00e9 limity (v\u00fd\u0161ka < hmax<\/sub>, n\u00e1klon < \\u03B8max<\/sub>, vzdialenos\u0165 z\u00f3ny \u013eud\u00ed > dsafe<\/sub>) a ako penaliz\u00e1cie v odmen\u00e1ch. Pre safe RL<\/em> je u\u017eito\u010dn\u00e1 formul\u00e1cia CMDP s Lagrangeov\u00fdm relaxa\u010dn\u00fdm \u010dlenom pre poru\u0161enia.<\/p>\n Kritick\u00e1 je robustn\u00e1 lokaliz\u00e1cia platformy v re\u00e1lnom \u010dase:<\/p>\n Odmena by mala vies\u0165 k bezpe\u010dn\u00e9mu a hladk\u00e9mu prist\u00e1tiu bez lok\u00e1lnych optimum:<\/p>\n Kontinu\u00e1lny ak\u010dn\u00fd priestor favorizuje off-policy met\u00f3dy so stabilnou konvergenciou:<\/p>\n Prislab\u00fd prieskum v zlo\u017eitom priestore rie\u0161i postupn\u00e9 zvy\u0161ovanie n\u00e1ro\u010dnosti:<\/p>\n Preklenutie sparse reward<\/em> a ur\u00fdchlenie tr\u00e9ningu:<\/p>\n Verizmus kontaktu je k\u013e\u00fa\u010dov\u00fd pre bezpe\u010dn\u00fd prenos do reality:<\/p>\n Strat\u00e9gie zni\u017eovania medzery medzi simul\u00e1ciou a realitou:<\/p>\n RL politika mus\u00ed by\u0165 obalen\u00e1 mechanizmami s preuk\u00e1zate\u013en\u00fdm spr\u00e1van\u00edm:<\/p>\n Pri flotile UAV alebo viacer\u00fdch mobiln\u00fdch cie\u013eoch je potrebn\u00e1 koordin\u00e1cia:<\/p>\n Praktick\u00e1 integr\u00e1cia vy\u017eaduje jasn\u00e9 rozhrania:<\/p>\n Objekt\u00edvne porovnanie polit\u00edk je z\u00e1sadn\u00e9:<\/p>\n Scen\u00e1r: kvadrokopt\u00e9ra prist\u00e1va na 1.5\u00d71.5 m plo\u0161ine plavidla s bo\u010dn\u00fdm zr\u00fdchlen\u00edm do 0.3 g a n\u00e1hodn\u00fdm k\u00fdvan\u00edm (roll\/pitch \u00b16\u00b0). Politika SAC je tr\u00e9novan\u00e1 s curriculum od stacion\u00e1rnej paluby po pln\u00e9 spektrum v\u013an, s dom\u00e9novou randomiz\u00e1ciou hmotnosti \u00b115 % a odhadovan\u00e9ho oneskorenia kamery 60\u2013120 ms. Bezpe\u010dnostn\u00fd shield<\/em> s CBF garantuje v\u00fd\u0161ku > 0.5 m pri later\u00e1lnej chybe > 0.8 m a limituje n\u00e1klon na 20\u00b0. V ter\u00e9nnych sk\u00fa\u0161kach je dosiahnut\u00e1 92 % miera \u00faspe\u0161n\u00fdch prist\u00e1t\u00ed, medi\u00e1n later\u00e1lnej chyby 12 cm a maxim\u00e1lne vertik\u00e1lne pre\u0165a\u017eenie 1.8 g, pri\u010dom pri sea state<\/em> > 3 politika automaticky vol\u00ed go-around<\/em>.<\/p>\n \u00daspe\u0161n\u00e9 prist\u00e1tie nekon\u010d\u00ed dotykom. Mechanick\u00e9 rie\u0161enia (magnetick\u00e9\/akt\u00edvne uchytenie, h\u00e1k s vodiacimi hranami, pas\u00edvne tlmenie) v\u00fdrazne roz\u0161iruj\u00fa pracovn\u00fa ob\u00e1lku. RL politika m\u00f4\u017ee ako s\u00fa\u010das\u0165 akci\u00ed ovl\u00e1da\u0165 stav z\u00e1mku (lock\/unlock<\/em>) a v\u00fd\u0161ku dotyku s oh\u013eadom na citlivos\u0165 sp\u00fa\u0161\u0165a\u010da.<\/p>\n Prist\u00e1tia na pohybuj\u00face sa platformy \u010dasto prebiehaj\u00fa v zlo\u017eit\u00fdch priestoroch a v bl\u00edzkosti os\u00f4b. Nevyhnutn\u00e9 s\u00fa SOP, geofencing, detekcia os\u00f4b v bezpe\u010dnostnom perimetri, lost-link<\/em> postupy a logovanie pre audit. Pre kritick\u00e9 misie je vhodn\u00e9 form\u00e1lne pos\u00fadi\u0165 riziko (FMEA\/FTA) a zavies\u0165 dvojkan\u00e1lov\u00e9 dohliadanie politiky (runtime assurance<\/em>).<\/p>\n U\u010denie zosil\u0148ovan\u00edm pos\u00fava hranice auton\u00f3mneho prist\u00e1vania na pohybuj\u00face sa platformy t\u00fdm, \u017ee priamo optimalizuje spr\u00e1vanie v komplexn\u00fdch a neist\u00fdch podmienkach. K\u013e\u00fa\u010dom k nasadeniu je kombin\u00e1cia: siln\u00e1 percepcia a odhad, starostlivo navrhnut\u00e1 odmena, curriculum a dom\u00e9nov\u00e1 randomiz\u00e1cia, bezpe\u010dnostn\u00e9 shieldy<\/em> a hybridn\u00e9 riadenie s tradi\u010dn\u00fdmi regul\u00e1tormi. S rast\u00facou zrelos\u0165ou model-based RL, offline RL a multiagentn\u00fdch techn\u00edk mo\u017eno o\u010dak\u00e1va\u0165 \u0161ir\u0161ie nasadenie v logistike, n\u00e1mornej prev\u00e1dzke i v rojn\u00fdch scen\u00e1roch, kde sa auton\u00f3mne syst\u00e9my bud\u00fa spo\u013eahlivo prip\u00e1ja\u0165 k mobiln\u00fdm infra\u0161trukt\u00faram v re\u00e1lnom svete.<\/p>\n","protected":false},"excerpt":{"rendered":" U\u010denie zosil\u0148ovan\u00edm pre prist\u00e1tie na pohyblivej platforme: od simul\u00e1cie po bezpe\u010dn\u00fd transfer do re\u00e1lneho sveta.<\/p>\n","protected":false},"author":44,"featured_media":88691,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2629],"tags":[161,2512,2513,2180,2136,2514,2515,2516],"class_list":["post-48691","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-drony","tag-bezpecnost","tag-reward-shaping","tag-rl-pristatie-na-pohybe","tag-robustnost","tag-senzory","tag-simulator","tag-transfer","tag-zosilnovanie"],"yoast_head":"\nForm\u00e1lne vymedzenie \u00falohy a model prostredia<\/h2>\n
\n
Priestor stavov a akci\u00ed, bezpe\u010dnostn\u00e9 obmedzenia<\/h2>\n
Percepcia cie\u013ea a odhad stavu<\/h2>\n
\n
Kon\u0161trukcia odmeny a tvarovanie (reward shaping)<\/h2>\n
\n
Algoritmy RL: v\u00fdber a porovnanie<\/h2>\n
\n
Kurikul\u00e1rny tr\u00e9ning a dom\u00e9nov\u00e1 randomiz\u00e1cia<\/h2>\n
\n
U\u010denie z demon\u0161tr\u00e1ci\u00ed a offline RL<\/h2>\n
\n
Simul\u00e1tor, kontaktn\u00e1 fyzika a valid\u00e1cia<\/h2>\n
\n
Sim2Real: prenos politiky do praxe<\/h2>\n
\n
Bezpe\u010dnos\u0165 a form\u00e1lne garancie<\/h2>\n
\n
Multiagentn\u00e9 scen\u00e1re a roje<\/h2>\n
\n
Integr\u00e1cia s autopilotom a architekt\u00fara softv\u00e9ru<\/h2>\n
\n
Metodika hodnotenia a metriky<\/h2>\n
\n
Pr\u00edpadov\u00e1 \u0161t\u00fadia: prist\u00e1tie na palube plavidla<\/h2>\n
Hardv\u00e9rov\u00e9 aspekty: podvozok a zabezpe\u010denie kontaktu<\/h2>\n
Prev\u00e1dzkov\u00e9 a regula\u010dn\u00e9 h\u013eadisk\u00e1<\/h2>\n
Best practices pre implement\u00e1ciu<\/h2>\n
\n