Game AI

Responsible: Ben Goertzel (architecture); Berick Cook (AIRIS integration)

Papers: Hyperon for AGI⇒ASI Whitepaper (2025), §9.1

GitHub: Vereya (Minecraft mod), minecraft-demo (Python API), minecraft-experiments, AIRIS-client, rocca (Rational OpenCog Controlled Agent — OpenAI Gym RL), axiom (AXIOM object-centric game-world agent)

Status: Active pilot. Minecraft integration operational via Vereya mod and tagilmo Python API. AIRIS demonstrated causal learning in Minecraft without pre-training. Sophiaverse and Neoterics playground are under development.

Games provide ideal sandboxes for testing perception, planning, tool use, dialogue, and social norms with rapid iteration and safe failure. Three environments are targeted:

• Minecraft — A structured yet open world for integrating SubRep options with evolutionary program edits. The Vereya Fabric mod (Java 21) exposes game state via a network API; the tagilmo Python library provides high-level agent control.
• Sophiaverse — Adds persistent identity, social contracts, and economic systems to the game AI substrate.
• Neoterics — A deliberately constrained micro-world within Sophiaverse, optimized for baby-AGI development with short feedback loops, dense sensor coverage, and cheap resets.

Technical Architecture

The whitepaper describes the game interface operating through Spaces that mirror game state (voxels, entities, quests, markets) into AtomSpace as typed Atoms. Action affordances appear as SubRep options (navigate, craft, trade, negotiate) with MeTTa rules expressing pre/post-conditions. External LLMs handle narration and quest generation initially, with Symbolic Heads retrieving mined templates like task frames and recipes. PLN factor-graphs encode precondition networks and multi-step plans, with geodesic control scoring candidate steps.

AIRIS (Autonomous Intelligent Reinforcement Inferred Symbolism) has demonstrated causal learning in Minecraft — constructing deterministic environment models through direct interaction without pre-training, and self-correcting via scientific method application.

Additional Repos

• rocca — Rational OpenCog Controlled Agent. Python RL agent using OpenCog AtomSpace for OpenAI Gym environments. Demonstrates planning via PLN and action selection via cognitive schematics. Legacy OpenCog era but conceptually aligned with the PRIMUS game-AI architecture.
• axiom — AXIOM (Adaptive Expansion Object-Centric Models). JAX-based system for rapidly learning to play video games through object-centric representation learning — discovers and tracks objects from raw pixels using hierarchical slot mixture modeling with MPPI-based planning. Developed by VERSES AI. Relevant as a modern object-centric game-world agent architecture, though not directly integrated with MeTTa or AtomSpace.

Key References

• Goertzel, B. (2025). Hyperon for AGI⇒ASI Whitepaper, §9.1: Game AI
• AIRIS — causal machine learning system