Game Worlds and Simulated Environments

Draft — This content has not been approved for publication.

Scope

Repositories connecting Hyperon/OpenCog agents to game worlds, simulated environments, and reinforcement-learning benchmarks. This family covers the Minecraft integration stack, autonomous agent architectures designed to learn and plan within simulated worlds, and object-centric RL research. Physical and software-domain robotics are covered in Robotics and Embodiment; the NARS reasoning libraries that some of these agents build on are in NARS Ecosystem.

Active Repositories

Minecraft Integration Stack

A layered pipeline: Java mod → Python API → agent clients. The core runtime (Vereya, minecraft-demo, minecraft-experiments) is maintained by TrueAGI; AIRIS-client is a SingularityNET project that bridges into this pipeline.

Repo	Language	Maturity	Purpose
Vereya	Java 21	Active	Minecraft 1.21 Fabric mod exposing a low-latency network API for agent control. Voxel grids, entity observations, XML mission definitions, advanced image segmentation. Foundation for all Python-based Minecraft agents.
minecraft-demo (tagilmo)	Python 3.10+	Stable	Canonical Python client for Vereya. Published as the `tagilmo` package. Provides VereyaPython client, mission builder, environment wrapper, and integration test suite.
minecraft-experiments	Python	Active Research	Research sandbox extending tagilmo — And-Or behavior trees, visual perception tests, voxel feature collection, micro-skill training, RL experiments, and vision-language-action studies.
AIRIS-client	Python 3.10	Stable	Bridge connecting AIRIS agents to Minecraft via tagilmo. Dual session backends: local HTTP API and Fetch uagents protocol. Exports session state as NumPy artifacts for offline analysis.

Autonomous Agent Architectures

Four distinct approaches to learning and planning in simulated environments — from symbolic causal reasoning to deep RL with object models.

Repo	Language	Maturity	Purpose
AIRIS_Public	Python (pygame, numpy, numba)	Active	Autonomous Intelligent Reinforcement Inferred Symbolism — a general machine intelligence system combining RL with symbolic rule learning and planning. Puzzle games, CartPole, MountainCar demos. Visual planning via minds_eye. Knowledge persistence via NumPy serialization.
AIRIS-general	Python (stdlib only)	Research Library	Lightweight single-file (~78 KB) AIRIS library for causal discovery. Zero external dependencies. Oracle prediction mode or goal-directed planning mode. Designed for embedding into arbitrary environment wrappers.
NACE	Python (matplotlib)	Active Research	Non-Axiomatic Causal Explorer — extends AIRIS concepts with NAL truth values, partial observability, and non-deterministic environment handling. 20+ built-in worlds (custom grids, MiniGrid/gymnasium). Cognitive schematics: (precondition, operation) ⇒ consequence. Optional MeTTa and ONA integration. Reported 1000× more data-efficient than deep RL on comparable tasks.
rocca	Python 3.10 (nbdev)	Active Research	Rational OpenCog Controlled Agent — symbolic reasoning over OpenAI Gym environments using PLN, pattern mining, and Thompson Sampling. Requires the full OpenCog C++ stack (cogutil, atomspace, unify, ure, spacetime, pln, miner). Minecraft examples via MineRL (not Vereya). Docker dev container with VNC available.
axiom	Python 3.10–3.11 (JAX, Equinox)	Very Recent	Adaptive Expansion Object-Centric Models — four-model system (Slot Mixture, Temporal, Identity, Reward/Relation) with MPPI planning and Bayesian model reduction. Targets the Gameworld 10k benchmark. GPU-accelerated via JAX. Published paper.

How They Fit Together

This family spans two integration stacks and three independent agent architectures:

Minecraft pipeline: Vereya (Java mod) → minecraft-demo/tagilmo (Python API) → minecraft-experiments (research sandbox). This core pipeline is maintained by TrueAGI. AIRIS-client (SingularityNET) adds a bridge layer, connecting the AIRIS agent systems to this same Minecraft environment via tagilmo.

AIRIS family: AIRIS_Public is the full standalone system with GUI demos and multiple environments. AIRIS-general is a minimal embeddable library with the same core causal reasoning but zero dependencies. Either can be wrapped by AIRIS-client for Minecraft deployment. NACE extends the AIRIS approach with NAL-based truth values and partial observability, running its own grid world and gymnasium environments rather than Minecraft.

rocca (OpenCog symbolic): An independent agent framework built on the full OpenCog C++ stack (AtomSpace, PLN, URE, pattern miner). Uses Thompson Sampling for action selection and PLN for learning temporal patterns. Its Minecraft integration uses MineRL/Malmo rather than the Vereya pipeline — a separate stack that could theoretically be bridged to Vereya via tagilmo.

AXIOM (deep RL): Entirely independent of both integration stacks. Pure JAX-based object-centric world model with no symbolic reasoning component. Represents the neural end of the spectrum that Hyperon's neural-symbolic vision aims to bridge.

Quick Start

# Vereya (requires Java 21 + Minecraft 1.21 + Fabric Loader)
cd Vereya && ./gradlew build
cp ./build/libs/vereya-*.jar ~/.minecraft/mods/

# minecraft-demo / tagilmo (Python client)
pip install git+https://github.com/trueagi-io/minecraft-demo.git
cd tests/vereya && python run_tests.py

# AIRIS_Public (standalone demos)
cd AIRIS_Public && python puzzle_game_driver.py

# NACE (grid worlds + gymnasium)
cd NACE && python main.py                    # GUI
cd NACE && python main.py world=9 interactive  # MeTTa-NARS shell

# rocca (requires full OpenCog stack or Docker)
cd rocca && pip install -e . && jupyter notebook 01_cartpole.ipynb

# AXIOM (requires JAX)
cd axiom && pip install -e . && python main.py --game=Explode

Excluded from This Family

nartech_ros: ROS 2 robotics integration for NARS. Uses Gazebo simulation, but the focus is physical robot control rather than game-world agent learning. In Robotics and Embodiment.
metta-nars, OpenNARS-for-Applications, estream: NARS reasoning libraries that NACE and AIRIS build on. In NARS Ecosystem.
NarsGPT, NARS-GPT: GPT+NARS wrapper projects — application integrations, not agent architectures.
FabricPC: JAX predictive coding library — pure ML research with no game-world coupling.
NAC-Experiments: NGC framework replication — ML research, no agent/environment loop.

Current State vs. Whitepaper

Game AI (whitepaper §9.1): The whitepaper describes a Minecraft agent demonstrating "mine → smelt → craft → trade" multi-step planning. The Vereya/tagilmo stack provides the runtime, but no current agent achieves this full pipeline. NACE's causal schematics and rocca's PLN planning are the closest reasoning approaches.
GEO-EVO corridor: The whitepaper envisions geodesic-guided evolutionary learning for game agents. AXIOM's object-centric models and NACE's causal explorer represent partial approaches from opposite ends of the neural-symbolic spectrum, but neither implements the full GEO-EVO vision.
Data efficiency claims: NACE reports 1000× greater data efficiency than deep RL on comparable tasks, aligning with the whitepaper's motivation for symbolic reasoning in agent control.

Forks and Mirrors

No numbered mirror/fork pattern within this family. AIRIS_Public tracks berickcook/AIRIS_Public (original author). AIRIS-client and AIRIS-general are SingularityNET repos. NACE and AXIOM are patham9 repos. rocca is under the opencog organization.

Gaps and Consolidation Opportunities

No unified agent ↔ environment abstraction: Each agent system defines its own observation/action interface. A shared MeTTa-native environment wrapper would allow NACE, AIRIS, and rocca agents to target the same environments without per-system bridging code.
Two Minecraft stacks: Vereya/tagilmo (active, Minecraft 1.21) and MineRL/Malmo (rocca, older Minecraft versions). Converging on Vereya would consolidate maintenance and enable agent comparison on a shared platform.
NACE ↔ MeTTa integration is optional: NACE supports MeTTa and ONA integration but doesn't require it. Deeper MeTTa coupling would make NACE's causal schematics accessible to the broader Hyperon reasoning stack.
AXIOM has no symbolic bridge: Pure neural approach with no MeTTa or AtomSpace integration. Represents an opportunity for neural-symbolic fusion — object models from AXIOM feeding symbolic reasoning in NACE or PLN.
No MORK-native agent: All agent architectures operate above the substrate layer. MORK's trie structures could potentially accelerate NACE's knowledge base or rocca's pattern mining.