Okolja zavedajoči z LLM-ji podprti neigralni liki v igrah

Radež, Grega (author), Bohak, Ciril (mentor)

Diplomska naloga predstavlja nov sistem uporabe LLM-jev, ki izboljšuje delovanje NPC-jev v igrah in jim omogoča kontekstualno zavedanje svojih okoliščin. Tovrstna rešitev ponuja dinamične interakcije, ki so povezane z okoljem. Tradicionalno so NPC-ji odvisni od vnaprej napisanih dialogov in jim primanjkuje zavedanje svojega okolja, kar omejuje pristnost njihovega odziva na igralca. V nalogi smo se te težave lotili z zajemanjem panoramične slike NPC-jeve okolice. Ta slika je nato uporabljena kot vhod za semantično segmentacijo, ki služi identifikaciji objektov. Generiramo strukturirano JSON predstavitev NPC-jevega okolja s kombinacijo lokacij objektov ter informacij pridobljenih iz segmentacije. Ti podatki so potem podani kot kontekst LLM-ju, kar NPC-ju omogoči, da jih uporabi v pogovoru z igralcem. Rezultat je bolj poglobljena igra, kjer se lahko NPC nanaša na bližnje objekte in okoljske značilnosti, kar naj bi omogočilo pristnejšo izkušnjo v igri. Opisane so metode in tehnična implementacija sistema, ki prikaže, kako lahko integracija ”vida”v NPC-je preobrazi dialoge in interakcije z le-temi.

• Language: Slovenian
• Keywords: zavedanje okolja, dinamični pogovori, semantična segmentacija, kontekstualno obči NPC-ji, LLM, NPC, razvoj, igre
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-166075
• COBISS.SI-ID: 220611843
• Publication date in RUL: 19.12.2024

RADEŽ, Grega, 2024, Okolja zavedajoči z LLM-ji podprti neigralni liki v igrah [online]. Bachelor’s thesis. [Accessed 23 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=166075

Secondary language

• Language: English
• Title: Environment-aware LLM-supported non-playable characters in games

Abstract

The thesis presents a novel system that enhances NPCs utilizing an LLM in games with contextual awareness of their surroundings, offering dynamic, environment-sensitive interactions. Traditionally, NPCs rely on pre-scripted dialogue and lack awareness of their environment, limiting their responsiveness to player actions. Our system addresses this by capturing a panoramic image of the NPC’s surroundings and applying semantic segmentation to identify objects and their spatial positions. We generate a structured JSON representation of the NPC’s environment by combining object locations with segmentation information. This data is provided as context to a LLM, enabling NPCs to incorporate spatial knowledge into their conversations with players. The result is more immersive gameplay, where the NPC can reference nearby objects, landmarks, and environmental features during interactions, enhancing believability and engagement. The thesis discusses the technical implementation of this system, demonstrating how integrating visual perception into NPCs can transform in-game dialogues and interactions.

• Keywords: context-aware NPCs, semantic segmentation, immersive gameplay, environmental perception, dynamic interaction systems, LLM, NPC, development, games