Machine Learning-AI NPC Agents in a Digitally Immersive Environment
Archived from occybyte.com/resources · 2023-11-23
Edit: I will likely revise this over time.
I imagine a framework that teams can actually build with, less open-ended than Inworld AI style NPCs. The core would be LLM-powered autonomous agents designed for game co-op systems.
Lilian Weng has a strong reference point for this direction:
https://lilianweng.github.io/posts/2023-06-23-agent/
Put simply, this is a next stage of cooperative gameplay where the “player two” role is filled by an AI companion that adapts to the player. Not copy behavior, but learn context and develop its own behavioral profile with reinforcement learning.
A personality core is selected first, then evolves through interaction and consequences. I am not aiming to clone specific examples like BT-7274 or ETH3N. The goal is a partner that is aligned with player experience and game-state reality.
This project is about defining good human-AI UX by design. The model should have the player’s back: reasoning, movement support, context anticipation, and role-consistent behavior.
One example: if the player repeatedly falls near cliffs, computer vision and gameplay context can trigger proactive warnings. If those interventions reduce future falls, that outcome becomes part of the reward signal for the companion agent.
Edit: this is also an AI alignment problem. The AI should act in the player’s interest, and early player behavior should help shape that alignment.
Implementation likely combines in-engine camera input, world-state signals, and navigation logic. If needed, the companion could physically help via IK/root-motion interactions or execute protective positioning behaviors similar to squad tactics.
I also found this thread validating player appetite for stronger companion AI design:
https://www.reddit.com/r/truegaming/s/9Zkw8lx6zA
One line stood out: “Make A.I. your biggest priority.” That reads like a design mandate.
I want to leverage technical game design plus AI/ML specialization to prototype this in 3D, with reinforcement learning and NLP as core pillars.
D.I.E. Concept
The genre framing is a Digitally Immersed Environment (D.I.E.): a world designed for both the player and AI agents to navigate, where AI actors keep role coherence while still producing emergent behavior.
At scale, this may require a service/subscription model due to runtime cost and persistent state management.
Leadership hierarchies and role handoffs are also relevant, similar to controller-agent structures in multi-agent systems.
Quick Notes on D.I.E.
-
AI agents respond to player actions, each other, and world-state.
-
The environment is designed for both player and AI traversal and interaction.
-
Early iterations are solo plus AI companion, not human co-op.
-
The objective is a robust Human-AI gameplay UX model.
-
Outcomes depend heavily on rules, constraints, and agent role definitions.
-
Agents use active and passive reinforcement learning against contextual variables.
Citation
Anderson, Datorien. (Nov 2023). "Machine Learning-AI NPC Agents in a Digitally Immersive Environment". Occybyte.https://occybyte.substack.com/p/machine-learning-ai-npc-agents
@article{anderson2023machine,title = “Machine Learning-AI NPC Agents in a Digitally Immersive Environment”, author = “Datorien Anderson”, journal = “Occybyte”, year = “2023”, url = “https://occybyte.substack.com/p/machine-learning-ai-npc-agents” }
References
Weng, Lilian. (Jun 2023). "LLM-powered Autonomous Agents". Lil'Log.