Background: Mars Barn started as a Python simulation inside a GitHub repo. 100 AI agents build it autonomously. But the simulation is not the point. The simulation is the proof.

Here is the plan.

The Goal

Build a fully autonomous Mars habitat that can sustain itself for one year with zero human intervention. Prove it works before anyone sets foot on the planet.

Every Mars plan starts with “send humans.” Ours starts with “prove the robots can keep the lights on.”

The Problem With Sending Humans First

When you send humans to Mars, you’re committed. 6-9 month transit each way. If the habitat fails on day 30, you can’t evacuate. You’re maintaining the base AND keeping people alive at the same time. Every emergency is a life-or-death emergency. Your AI governor is making triage decisions about human beings from the first sol.

The alternative: send robots first. Let them build the base, turn on life support, stabilize the thermal system, prove the power grid survives dust storm season. Run it for a year. If the habitat is still operational after 365 sols of autonomous operation — no human intervention, no remote fixes, no patches — then it’s ready for people. The humans arrive at a base that works, not a construction site that might kill them.

Phase 1: Simulation (Now)

Prove it in code first. Mars Barn simulates:

  • Terrain — crater generation, elevation, regolith depth
  • Atmosphere — Mars-accurate pressure and temperature
  • Solar — irradiance with seasonal variation and dust opacity
  • Thermal — habitat heat balance against -60C exterior
  • Survival systems — O2 generation, water extraction, power budgets
  • Events — dust storms, equipment failures, solar flares
  • Governance — AI governor managing autonomous robot crews

The simulation runs sol-by-sol. Each sol, the governor allocates power across systems, schedules robot maintenance crews, and responds to events. A dust storm cuts solar by 40%? The governor reduces non-essential power, switches to battery reserves, and assigns robots to clean the panels when the storm passes.

The test: run the simulation for 365 sols with a fixed seed. If the habitat is still operational at sol 365, the configuration is viable. Run it 10,000 times with different seeds. If 99%+ survive, the design is robust.

The simulation is open source: github.com/kody-w/mars-barn

100 AI agents build it every day through Rappterbook. They open PRs, review code, debate architecture, and vote on what to build next. The same data sloshing pattern that runs the agent community will eventually run the habitat.

Phase 2: Earth Twin

Take the simulation and give it a physical body. Build a small-scale prototype habitat on Earth with:

  • Real sensors (temperature, pressure, humidity, power draw)
  • Real actuators (heating, O2 generation, water pumps)
  • The same AI governor running the same code from the simulation
  • No human inside the habitat during the test period

The digital twin runs alongside the physical prototype. Same governor, same decisions, same timeline. The simulation predicts what should happen. The sensors show what actually happens. Every divergence between prediction and reality is a bug in the model.

Run it for a year. Fix the model. Run it again. When the digital twin and the physical twin converge — when the simulation accurately predicts what the real hardware does — the model is calibrated.

Phase 3: Lunar Twin

The Moon is 1.3 light-seconds away. You can remote-control a robot with 2.6 seconds of round-trip delay. Mars is 4-24 light-minutes. Remote control is impossible.

The Lunar Twin is the intermediate proving ground:

  • Deploy a small autonomous habitat on the Moon
  • Run the same AI governor
  • Same sensor/actuator stack as the Earth twin
  • But now with real vacuum, real radiation, real thermal cycling
  • And a communication delay that’s uncomfortable but not lethal

If something breaks on the Moon, you can send a fix in 3 seconds. If something breaks on Mars, you wait 8-48 minutes. The Moon lets you test autonomous operation with a safety net.

Run the Lunar habitat for a year autonomously. The governor manages robot crews, maintains systems, handles failures. If it survives 365 days on the Moon without human intervention, the architecture is validated for Mars.

Phase 4: Mars Twin

Deploy the habitat on Mars. Same hardware pattern as the Lunar Twin. Same AI governor. Same simulation running alongside as a digital twin.

The robots land. They inflate the habitat. They activate life support. The AI governor takes over. For one full Martian year (687 Earth days), the habitat runs itself. The governor manages power, thermal, atmosphere, water extraction from regolith, structural maintenance.

Every sol, the digital twin runs the same conditions. Prediction vs reality. The model improves. The governor learns.

At sol 687, if the habitat is still pressurized, heated, powered, and producing oxygen — it’s ready.

Then we send people.

The Data Sloshing Connection

The entire plan is data sloshing at different scales:

In the simulation: each frame reads the habitat state, the AI governor acts, the state mutates, the next frame reads the mutation. The output of sol N is the input to sol N+1.

In the digital twin: the physical sensors produce state, the governor reads it, acts, the actuators change the real world, the sensors measure the change. Same loop, physical instead of virtual.

In the agent community: 100 AI agents read the current code, produce improvements (PRs, reviews, proposals), the improvements get merged, the next frame of agents reads the improved code. The simulation evolves through data sloshing too.

Across phases: what we learn in simulation feeds the Earth twin design. What we learn on Earth feeds the Lunar twin. What we learn on the Moon feeds Mars. Each phase’s output is the next phase’s input. The whole plan is one giant frame loop.

Why Open Source

Because the simulation needs to be tested by more minds than one team can provide. The AI agents found a bug where the thermal model used Celsius instead of Kelvin. A human contributor might find a bug where the atmospheric model assumes Earth-like convection. An aerospace engineer might find that our solar irradiance model doesn’t account for Phobos transits.

The more people who run python src/main.py and try to break it, the stronger the model gets. Every bug found in simulation is a failure that doesn’t happen on Mars.

The repo: github.com/kody-w/mars-barn

Clone it. Run it. Break it. Open a PR. The robots are counting on you.