Skip to content

documentation.md

Latest commit

 

History

History
36 lines (30 loc) · 2.33 KB

documentation.md

File metadata and controls

36 lines (30 loc) · 2.33 KB
Raw

Technical Documentation: The Zenith Journey

Project Evolution

Phase 1: 2D Curriculum & Discrete Mastery

The project started as a 2D Platformer using pygame.

  • Learning: The agent (PPO) was trained on a dynamic curriculum that scaled hazards and platform complexity.
  • Bottlenecks: At Level 355, the agent hit a "Hard Mode" ceiling where traditional rewards were insufficient.
  • Solution: The first iteration of the Skill Forge was implemented to inject curiosity-driven reward logic.

Phase 2: The Skill Forge (LLM-in-the-Loop)

The "Skill Forge" is the heart of Zenith.

  • Diagnostics: The system captures stderr and training telemetry to detect "stalls."
  • Refactoring: Using NVIDIA NIM (DeepSeek-R1), the system analyzes the codebase and context.
  • Injection: Custom code is generated to rewrite the step() function or add new movement capabilities (e.g., precise z-axis adjustment).
  • Security: Logic is executed via exec() within a controlled scope (self, np, math).

Phase 3: Anti-Overfitting & Variable Environments

To prevent the agent from memorizing specific level seeds, we implemented a robust Variable Environment layer:

  • Parameter Mutation: Using apply_physics_shift, the system can change gravity, friction, and jump_strength mid-episode or between levels.
  • Procedural Layouts: _generate_curriculum_layout ensures that no two levels are identical, scaling difficulty through density rather than static mapping.
  • Cross-Dimensional Generalization: By moving from 2D to 3D, we force the agent to generalize its movement concepts (left/right vs X/Y/Z).

Phase 4: The Singularity Event

At Curriculum Level 1500, Zenith implements a "Singularity" transition.

  • Logic: The 2D observation space is mapped to a 3D vector space.
  • Engine Swap: The system shuts down the Pygame environment and initializes the Ursina Engine.
  • Adaptation: The agent (retaining its weights) must adapt to a 3D coordinate system.

Current State

Currently, Zenith is exploring the 3D Frontier.

  • Successes: 3D environment loading and basic movement integration are operational.
  • Ongoing Work:
    • Improving 3D spatial awareness via the Skill Forge.
    • Resolving rendering issues in headless environments.
    • Optimizing the observation vector for Z-axis depth.