Outbreak Protocol

Bears
Notes from the LLM are sent back to discord so friends can add more context to the bot. The prompt to the LLM includes this additional info.
Output from LLM to highlight the actions it recommended to take and the reason to take them.
GIF
Example showing how the LLM adjusts to make things more challenging by spawning ammo behind the bears nearby the player.
GIF
While exploring an area with no pickups or enemies, the LLM spawned multiple enemies behind the player.
GIF
Initial spawn the LLM would create more bears than expected, creating more fun than I expected.

Inspiration

Low Ammo

The idea for Outbreak Protocol was born from a desire to explore and experiment with cutting-edge technologies in UE5 and AWS. While the gameplay is a simple zombie FPS shooter—with the twist of zombie bears—it serves as a sandbox for testing and learning.

Key Learning Goals

Low Ammo

Unreal Engine 5 Exploration
- It allowed me to deepen my understanding of UE5’s tools for building engaging and reactive gameplay experiences.
AWS Integration
- The project utilizes several AWS offerings to enable scalability, real-time interactions, and AI-driven gameplay.
- Specifically:
  - Amazon GameLift supports dedicated servers for seamless multiplayer experiences.
  - Amazon EKS hosts a custom bot that processes real-time input from external sources.
  - Amazon Bedrock powers the Retrieval-Augmented Generation (RAG) system, enabling natural language interactions.
AI and Context-Driven Gameplay
- The most unique feature is the use of Claude, invoked via API calls, to analyze the current game state and generate actions directly in-game.
- For example:
  - If the game state shows the player is low on health and surrounded, Claude may spawn additional ammo nearby.
  - When friends suggest chaotic ideas like “What if we spawn an explosive toilet in the meadow?” the bot translates this into direct in-game events.
- Creates emergent gameplay moments where AI actively participates in shaping the player's experience.

LLM Output

Challenges we ran into

Gameplay Example

A major challenge, that Amazon Q helped overcome, was tied to the amount of complexity in building the game. I used multiple Beta/Experimental features of UE5 and ended up hitting major issues with their implementation.

Accomplishments that we're proud of

That it worked ... I really didn't expect to see it all work together across Discord, EKS, BedRock, GameLift and UE5 Remote.

What we learned

Low Ammo

An LLM driving gameplay is way more fun than I realized! It reminds me of the AI Director of L4D, it had few options open to it for executing as actions but found ways to create new mechanics.