3D World Construction
- Floors, walls, ceilings, ramps
- Boxes and cover geometry
- Collision, gravity, friction
- Acceleration, air movement
- Jumping and falling states
cs2-ai-learning
by
Wyatt
Wyattjust base it off the text doc, and have the gui clean too
Comments (0)
Sign in to leave a comment
System Requirements
Page 1 of 7
System Requirements Document for cs2-ai-learning
Introduction
This document outlines the requirements for a major upgrade to the existing project, "Neuro Duel." Neuro Duel is a reinforcement-learning experiment where a blank-slate AI learns to play a simplified CS2-style 1v1 game entirely through trial and error. The upgrade involves transitioning from a 2D simulation to a real-time 3D first-person FPS simulation while preserving existing learning, statistics, progression, and training concepts.
System Overview
The upgraded system will feature a real-time 3D first-person simulation using a browser-compatible 3D engine such as Three.js/WebGL. The AI will perceive the world through its own first-person camera and internal physical state, while a human spectator can view a top-down representation. The AI must learn spatial awareness and enemy location from its first-person visual input and experience.
Product Interpretation and Delivery Boundary
The system will deliver a real-time 3D simulation where the AI interacts with the environment through its first-person perspective. The human spectator will have access to a top-down view, but this information will not be shared with the AI. The project will maintain a dark technical training-lab aesthetic, focusing on simulation speed and correctness over visual quality. The application will support persistent AI training with reliable save/load functionality.
Page 2 of 7
Page Content and Component Coverage
Landing
- Information/State: Introduction to the 3D reinforcement-learning simulation, its audience, and core AI and spectator experience.
- Primary Actions: Navigate to Sign Up or Login.
- Components: Overview of AI capabilities, simulation features, and spectator view.
Login
- Information/State: User authentication for returning users.
- Primary Actions: Enter credentials, submit login.
- Components: Username and password fields, login button.
Sign Up
- Information/State: User registration for new users.
- Primary Actions: Enter registration details, submit sign-up.
- Components: Registration form fields, sign-up button.
Training Dashboard
- Information/State: Detailed training statistics, Policy Network activity, persistent brain controls, progression status.
- Primary Actions: Monitor AI progress, adjust training settings.
- Components: Statistics panels, Policy Network visualization, control buttons for Save Brain, Load Brain, Reset.
Page 3 of 7
Simulation Lab
- Information/State: Real-time 3D simulation workspace for AI vision, human-only top-down observation, physics, combat, progression, learning interaction, and training-speed controls.
- Primary Actions: Start/stop simulation, adjust training speed, observe AI and environment.
- Components: AI VISION panel, TOP-DOWN SPECTATOR view, training-speed controls, environment interaction elements.
Functional Requirements
Page 4 of 7
As a Developer:
- I want to replace the current 2D simulation with a real-time 3D first-person FPS simulation using Three.js/WebGL or another suitable browser-compatible 3D engine. (explicit)
- I want to ensure the AI perceives the world only through its first-person camera and internal physical state. (explicit)
- I want to create a genuinely 3D simulation with floors, walls, ceilings, ramps, boxes/cover, and simple structures. (explicit)
- I want to implement realistic physics including collision, gravity, friction, acceleration, air movement, player velocity, jumping, and falling. (explicit)
- I want to implement weapons with bullets/hitscan, damage, health, death, and respawning mechanics. (explicit)
- I want to give the AI a first-person camera for visual observation, which must be displayed in an "AI VISION" panel for the user. (explicit)
- I want the AI to know its own physical/game state, including health, ammo, current weapon, grounded/airborne status, and velocities. (explicit)
- I want to allow the AI to combine primitive actions such as movement, looking, and actions like firing and reloading. (explicit)
- I want to implement advanced movement physics that allow techniques like bunny hopping to emerge naturally. (explicit)
- I want to implement weapons including a knife, pistol, and rifle with specific mechanics. (explicit)
- I want to create six opponent tiers with increasing difficulty and preserve the existing promotion system. (explicit)
- I want to preserve the existing reinforcement-learning concept with first-person observations, actions, rewards, and punishments. (explicit)
- I want to implement a reward system with positive and negative rewards based on combat performance and survival. (explicit)
- I want to create a top-down spectator view for humans only, showing the 3D world layout and player positions. (explicit)
- I want to display both AI VISION and TOP-DOWN SPECTATOR views to ensure information separation. (explicit)
- I want to preserve and improve the existing Neuro Duel dashboard with detailed training statistics. (explicit)
- I want to create a visual "Policy Network" panel showing the AI's neural network activity. (explicit)
- I want to support various training speeds including PAUSE, PLAY, FAST, VERY FAST, and WATCH LIVE. (explicit)
- I want to implement persistent AI training with Save Brain, Load Brain, and Reset options. (explicit)
- I want to ensure the AI interacts with the 3D environment with real physics and weapon interactions. (explicit)
- I want to maintain a strict learning information boundary, ensuring the AI only receives permitted observations and rewards. (explicit)
- I want to maintain the existing Neuro Duel dark technical training-lab aesthetic for the UI. (explicit)
- I want to prioritize implementation of core features like the real 3D world, FPS camera, and reinforcement-learning loop. (explicit)
Page 5 of 7
User Personas
-
AI Developer: Responsible for implementing and maintaining the AI's learning algorithms and ensuring the simulation environment is conducive to reinforcement learning. The AI Developer focuses on the reinforcement-learning loop, observation/action/reward mechanisms, and persistent brain controls.
-
Game Designer: Focuses on the design and mechanics of the 3D world, ensuring the AI's interactions are realistic and challenging. The Game Designer is responsible for the environment's physics, opponent tiers, and progression mechanics.
-
Human Spectator: Observes the AI's progress and performance through the top-down spectator view and AI VISION panel. The Human Spectator monitors the AI's learning progress and training statistics.
Core User Flows
-
AI Training Flow:
- The AI starts in a 3D environment with no prior knowledge.
- It perceives the world through its first-person camera.
- It performs actions based on its policy network.
- It receives rewards or punishments based on its actions.
- It updates its learning model and improves over time.
-
Human Observation Flow:
- The human spectator views the AI's first-person perspective in the AI VISION panel.
- The spectator also sees the top-down view of the environment.
- The spectator monitors the AI's progress and statistics on the training dashboard.
Visuals Colors and Theme
- Palette: Dark mode with background #000022, surface #111111, text #FFFFFF, primary #00FFFF, accent #FF00FF, muted #333333.
- Typography: Headings in Space Grotesk, body in Orbitron, with a 1.5 modular scale.
- Shape Language: Floating glass panels, radial layouts, thin luminous strokes.
- Layout: Full-bleed 3D scene with floating UI elements overlaying the environment.
Page 6 of 7
Signature Design Concept
The design should evoke an experimental AI research laboratory, emphasizing the technical and experimental nature of the project. The interface will feature immersive 3D environments as heroes, with floating glass UI panels and data streams.
Interaction Model & Motion Direction
- Interaction Model: Animated
- Motion Tempo: Cinematic
- Hero Dimensionality: WebGL
- Landing Hero Motion Brief: The hero will feature a full-viewport 3D object set against a deep dark ground, with glowing elements and floating UI panels. The motion will include continuous slow orbit and parallax, with light sweeps and data streams.
Non-Functional Requirements
- The system must prioritize simulation speed and correctness over visual quality.
- The AI's learning and interaction with the environment must be real and not simulated through static mockups.
Tech Stack
- Use Three.js/WebGL for the 3D simulation.
- Implement reinforcement learning using a practical browser-compatible RL approximation.
Assumptions and Constraints
- The AI must not receive any information from the top-down spectator view.
- The system must support persistent AI training with reliable save/load functionality.
Page 7 of 7
Glossary
- AI VISION: The first-person view received by the AI.
- Top-Down Spectator: The human-only view of the 3D environment.
- Policy Network: The neural network used by the AI to determine actions based on observations.
- Reinforcement Learning: A type of machine learning where an agent learns by interacting with its environment and receiving feedback.
Preview dataChanges stay in this preview.
Neuro Duel
3D Reinforcement
Learning
A blank-slate AI learns a simplified CS2-style 1v1 game entirely through trial and error — perceiving the arena only through its own first-person camera and internal physical state, no human knowledge injected.
Observe→Action→Reward
First-Person Camera
Live AI-eye render feeds the policy network every simulated frame.
Learning Loop
Observation → Action → Reward, updated continuously as the AI trains inside the arena.
AI Vision
First-person only — no spectator data reaches the agent.
Training Arena
First-person camera, learning loop, and AI vision run inside this live chamber.
Reinforcement Learning Loop
First-Person Learning Loop
The AI begins with no prior knowledge of the arena. Every behavior it ever develops is built entirely from what it can perceive of itself and act on inside its own first-person circuit.
PolicyCore
- Visual ObservationFirst-person camera frame — the only visual signal the AI receives, rendered from its own eyes inside the arena.
- Physical StateHealth, ammo, current weapon, grounded/airborne status, and velocity — the AI’s internal body sense.
- Primitive ActionsMove, look, fire, and reload combine into every behavior the policy can express — nothing more.
- Reward / PunishmentCombat performance and survival generate positive and negative signal that reshapes the policy over time.
- Visual ObservationFirst-person camera frame — the only visual signal the AI receives, rendered from its own eyes inside the arena.
- Physical StateHealth, ammo, current weapon, grounded/airborne status, and velocity — the AI’s internal body sense.
- Primitive ActionsMove, look, fire, and reload combine into every behavior the policy can express — nothing more.
- Reward / PunishmentCombat performance and survival generate positive and negative signal that reshapes the policy over time.
Observation Architecture
Two separate views of the same arena. AI VISION is the only signal the agent ever receives. TOP-DOWN SPECTATOR exists solely for human observers and is never exposed to the learning loop.
Player positions and world layout are quarantined from the AI — visible to human spectators only.
Separate signals, separate panels — the void between them stays visible.
Simulation Systems
The training arena is a genuine 3D environment, not a scripted approximation — every surface, impact, and opponent tier below runs on real physics and real combat resolution.
Weapons & Damage
- Knife, pistol, rifle
- Bullet / hitscan resolution
- Health and damage tracking
- Death and respawning
Opponent Tiers
- Six opponent tiers
- Increasing difficulty per tier
- Preserved promotion system
Lab Priority
Simulation speed and correctness take priority over visual fidelity — every system above exists to keep the AI's first-person experience physically consistent, trial after trial.
Preview dataChanges stay in this preview.
Neuro Duel
3D Reinforcement
Learning
A blank-slate AI learns a simplified CS2-style 1v1 game entirely through trial and error — perceiving the arena only through its own first-person camera and internal physical state, no human knowledge injected.
Observe→Action→Reward
First-Person Camera
Live AI-eye render feeds the policy network every simulated frame.
Learning Loop
Observation → Action → Reward, updated continuously as the AI trains inside the arena.
AI Vision
First-person only — no spectator data reaches the agent.
Training Arena
First-person camera, learning loop, and AI vision run inside this live chamber.
Reinforcement Learning Loop
First-Person Learning Loop
The AI begins with no prior knowledge of the arena. Every behavior it ever develops is built entirely from what it can perceive of itself and act on inside its own first-person circuit.
PolicyCore
- Visual ObservationFirst-person camera frame — the only visual signal the AI receives, rendered from its own eyes inside the arena.
- Physical StateHealth, ammo, current weapon, grounded/airborne status, and velocity — the AI’s internal body sense.
- Primitive ActionsMove, look, fire, and reload combine into every behavior the policy can express — nothing more.
- Reward / PunishmentCombat performance and survival generate positive and negative signal that reshapes the policy over time.
- Visual ObservationFirst-person camera frame — the only visual signal the AI receives, rendered from its own eyes inside the arena.
- Physical StateHealth, ammo, current weapon, grounded/airborne status, and velocity — the AI’s internal body sense.
- Primitive ActionsMove, look, fire, and reload combine into every behavior the policy can express — nothing more.
- Reward / PunishmentCombat performance and survival generate positive and negative signal that reshapes the policy over time.
Observation Architecture
Two separate views of the same arena. AI VISION is the only signal the agent ever receives. TOP-DOWN SPECTATOR exists solely for human observers and is never exposed to the learning loop.
Player positions and world layout are quarantined from the AI — visible to human spectators only.
Separate signals, separate panels — the void between them stays visible.
Simulation Systems
The training arena is a genuine 3D environment, not a scripted approximation — every surface, impact, and opponent tier below runs on real physics and real combat resolution.
3D World Construction
- Floors, walls, ceilings, ramps
- Boxes and cover geometry
- Collision, gravity, friction
- Acceleration, air movement
- Jumping and falling states
Weapons & Damage
- Knife, pistol, rifle
- Bullet / hitscan resolution
- Health and damage tracking
- Death and respawning
Opponent Tiers
- Six opponent tiers
- Increasing difficulty per tier
- Preserved promotion system
Lab Priority
Simulation speed and correctness take priority over visual fidelity — every system above exists to keep the AI's first-person experience physically consistent, trial after trial.
No comments yet. Be the first!