Simulation Frameworks Guide

Mind Guild paper companion · researched 7 November 2026

Simulation Frameworks Guide

A guide to eight simulation options for an agent built to test theories of consciousness. Compare environments, frameworks, and agent interfaces by what they let an agent sense and do. Switch theory lenses, review the demo videos, and work through the open decisions.

“implement an agent with sufficiently expressive states and actions to test theories of consciousness”Minda, stating the goal
8names5worlds1genre1sandbox1agent9theory lenses33videos ranked12open decisions
nativepartial or buildableabsent or awkwardunknowninferredmy inference, not sourced↗source

Read this first

The eight names are four kinds of thing

Five are worlds, one is a genre, one is a sandbox built on another, and one is an agent. A comparison table that treats them as peers mixes levels.

FoundationResearch interface or platformWhat sits on top
MinecraftworldA Java game.
MalmoMineRLMineDojoMineStudioMineflayer
Each exposes a different observation and action space.
VoyageragentA language-model agent that reaches Minecraft through Mineflayer, reading text and writing code.
NetHackworldA 1987 terminal game.
NetHack Learning Environment (NLE)
The research interface to the full game.
MiniHacksandboxWrites custom levels on top of NLE.
MuJoCoworldMulti-Joint dynamics with Contact (MuJoCo), a physics engine with no world until you author one.
dm_controlGymnasiumMuJoCo PlaygroundMIMo body
Task suites and ready-made bodies.
Mind Guild scene familyin developmentListed on the Mind Guild framework page.
UnityA game engine.
ThreeDWorld (TDW)worldA Unity build driven by a Python controller.
nothing named
Godot + BulletGame and physics engines.
AvalonworldA simulator plus a 20-task benchmark.
nothing named
GridWorldgenreA family of environments, not one package.
MiniGridAI Safety GridworldspycolabJAX ports
The matrix uses MiniGrid as the reference implementation.
your choice of implementation

Voyager is an agent

Comparing Voyager with MuJoCo compares an agent with a physics engine. Voyager reaches Minecraft through the Mineflayer bot library.

“Minecraft” needs its interface named

MineRL gives the agent pixels and keyboard-and-mouse control. Mineflayer gives text and code. They sit at opposite ends of the interface range.

“GridWorld” needs an implementation named

MiniGrid, DeepMind’s AI Safety Gridworlds and the JAX ports differ in observations, speed and upkeep.

NetHack and MiniHack share an engine

They differ in experimental control. The full game generates its own levels; MiniHack lets you write them.

Compare

Affordance matrix

Pick a theory lens. The columns that lens depends on stay bright and the tally counts only those. Select any cell or column heading to see the fact behind it.

Order
Inspector

Select a cell or a column heading to see the fact and its source. Cells marked with a small red dot rest partly on my inference.

Specimens

Eight simulation options, side by side

Each card pairs what the agent receives with what it can emit, then lists what the world affords, what it strains, and facts that pull against each other.

Watch

Demo videos, ranked by what one viewing settles

For each environment, the question is whether a single video can tell you if the world can host the agent. Ranks favor videos that answer that fastest.

Not watched

I have not watched these videos. Both browsers disconnected before I could pull transcripts, so each rank rests on the title, channel, length, view count, upload age and the paper behind the video. The confidence label says how far that basis supports the rank.

A video ranks higher when it

  • shows the agent’s raw input, the frame or tensor it actually receives
  • shows the action interface, graded or discrete
  • shows a scene being built or changed in code
  • comes from the platform’s creators
  • shows an agent learning or acting
  • does all this in fewer minutes

The watch log under each video stays in your browser only. Use it to record what each video actually showed, so the next pass can skip what is settled.

Decide

Decisions this comparison forces

None are made here. Each lists facts that pull one way, facts that pull the other, a fact in tension, and what would settle it.

Vocabulary check

Words that shift meaning between fields

The paper spans Reinforcement Learning (RL), consciousness science and philosophy of mind. These terms read differently in each.

TermIn Machine Learning and RLIn consciousness science or philosophySuggested handling

Self-check

Where this page is reaching

Adjacent

Outside your list

Not researched in depth. Each touches a decision above.

Reference

Acronyms

Reference

Sources