NASA Space Apps Challenge 2026

Train for the Moon by talking to CAPCOM.

A voice-first desktop app. Learn in Flight School, then run an outpost at the lunar south pole. Every model runs on your computer.

Setup: about 770 MB in all. The app is 258 MB (Windows 194 MB); the AI models add 510 MB on first launch. Then it runs offline.
Runs on: any Apple silicon Mac (M1, 2020, or newer) with macOS 13 or later · Windows 10/11, 64-bit.

Free · No account · All releases

The Moon in space with a glowing marker at Shackleton Rim, near the south pole: the site of Outpost Luna.
Moon: NASA/GSFC/Arizona State University (LRO), via the NASA SVS CGI Moon Kit.

The app

Learn it, run it, explore it.

Artemis II flew around the Moon in April 2026. You train to run a south-pole base for the crews who come next.

  1. A Flight School lesson beside the Moon: Why does the Moon seem to change shape each month? Three answers to pick from.

    Flight School

    Spoken lessons with CAPCOM and real Artemis II crew audio.

    6 core lessons · 2 explorer lessons · a check ride

  2. The mission cabin at Shackleton Crater Rim during a solar particle storm alert, with power, life support and dose gauges.

    Mission

    Run the outpost for 10 days through real solar storms.

    5 south-pole sites · real Sun and Earth · debrief

  3. The Moon Library: a 3D Moon with hardware markers and a card for the Apollo 11 lunar module descent stage, with its NASA sources.

    Moon Library

    Hardware left on the Moon, each piece with its sources.

    36 pieces · 3D Moon · Artemis II crew moments

  • Play your way
  • Voice
  • Tap
  • Type
  • Keyboard
  • Game controller
  • Hand gestures

See every screen in the guide →

NASA open data

Real missions, real Moon.

Every lesson and event traces to a NASA source.

  • 74Artemis II crew clipsNASA Artemis audio
  • 13Real solar storms, 2017–2025NASA DONKI
  • 5South-pole sites, scoredLRO sunlight, Earth view, ice and slope
  • 17NASA videos, 8 play offlineNASA SVS and NASA channels
  • 36Pieces of hardware on the MoonNASA NSSDCA
  • 10Days of hourly Sun and EarthJPL Horizons, per site

Built into the app. No NASA calls while it runs.

How it works

Voice AI that runs on your computer.

Our own speech and language models hear, check and decide each turn. No cloud.

~13%word errors, Artemis II crew audioNVIDIA base model: 29%
7.0%word errors, NASA podcastsNVIDIA base model: 9.7%
~95%names linked correctly, Artemis audioThe base model can't link names

Our on-device ear (Arc Transcriber Space) vs the off-the-shelf NVIDIA model it builds on, on held-out NASA audio. Full results

  • Hears you, not the roomKnows when you're talking to the app.
  • Keeps it cleanSwearing is masked, never graded or answered.
  • Decides by rulesClear rules pick each move; a learned model trains alongside.
  • Speaks in one voice889 recorded lines, plus your computer's own voice.
Outpost Luna architecture On the computer: the microphone feeds the Arc ear (Arc Transcriber Space: live words, whether you are talking to the app, and turn end), then Arc Tagger v0.3 (safety masking, intent, places), then the Director (decides what happens next). The camera feeds MediaPipe hand and face models, whose gestures become the same intents. The Director picks a lane: voice out (the system voice plus the recorded Nora pack), facts and places (a MiniLM embedder over NASA facts), the mission engine (storms, power, ice) or the Flight School grader. NASA data from Horizons, DONKI, LROC, LOLA and NSSDCA is fetched at build time and built into the app. Online and optional: NASA video files, downloaded once then played offline, and YouTube embeds from NASA channels, which stream only. On your computer Online, only when available Micopen or hold Space Arc ear · Transcriber Spacewords · to the app? · turn end Arc Tagger v0.3safety mask · intent · places Cameraopt-in only MediaPipe hand + facegestures → the same intents Directorwho speaks · what happens next Voice outOS voice + Nora pack Facts & placesMiniLM over NASA facts Mission enginestorms · power · ice Flight Schoolgrader · check ride Lanes · act, ignore or navigate · all offline NASA data, built into the appHorizons Sun & Earth · DONKI storms · LROC / LOLA Moon maps · NSSDCA hardware sites NASA video filesdownloaded once, then offline YouTube embedsNASA channels · stream only NASA open-data APIsat build time, never at run time

Swipe to see the whole diagram.

Everything that answers works offline. Only the NASA YouTube clips need a connection.

How the ear is built and measured →

Privacy

Private by design.

No account. No cloud model. Nothing uploaded.

On-device
Every model runs on your computer.
Mic
Hands-free or push-to-talk. Audio is never stored.
Camera
Off until you turn on gestures.
No account
A call sign, kept on the device.
Turn log
Opt-in, text only. Never uploaded.
Privacy details
  • The app window may only reach its own local servers. Everything else is blocked, apart from the NASA YouTube player.
  • Hands-free listening is the default. Push-to-talk (hold Space) is in Settings.
  • While the camera runs, a "Camera on" chip with a Turn off button is always visible. Frames are never saved or sent. No face recognition.
  • Your call sign and all progress stay on the device.
  • The turn log ("Help improve the voice guide") is off by default, redacts names, and "Delete what's saved" erases it.

Team, challenge and credits

Team

Outpost Luna: Karan Singla and Bryan Pellom. We built the app and its on-device speech and language models.

Challenge

NASA Space Apps Challenge 2026: Build a Junior Astronaut Mission Trainer. Slides (PDF)

Credits

Moon: NASA/GSFC/ASU (LRO). Terrain: NASA LRO LOLA. Earth: NASA Blue Marble. Crew audio: NASA Artemis audio.

AI use

AI tools helped write code and draft lessons, and made the icon and some voices. The team checked every fact.

Machine-readable: ai-disclosure.json

AI use details
  • Code: written with AI coding assistants. Commits made with them carry a co-author line. The team reviewed and tested every change.
  • Writing: lesson text, "Why?" cards and CAPCOM lines were drafted with AI help. Every fact is checked against the NASA source it cites.
  • Data: part of the speech model's training data is synthetic text-to-speech audio. Results are measured on held-out clips; the one synthetic test set is labelled as such.
  • Images: the app icon and the team image were made with an AI image generator, with no NASA logos, flags or insignia. Moon and Earth imagery is NASA's.
  • Audio: CAPCOM's 889 scripted lines were recorded from macOS's built-in Siri voice, which is computer-generated. Live replies use the computer's own voice. The optional Steer voice is AI text-to-speech. The launch film's learner lines are AI-generated speech.
  • Not AI: NASA data, NASA video and crew audio, and app screenshots.