Autonomous Reasoning & Integration Architecture — a safety-first onboard AI for spacecraft. One reasoning engine plans the mission; a tamper-proof "constitution" and an independent monitor decide what it's actually allowed to do.
Source-available · research prototype (TRL 3–5) · not flight-qualified
The ARIA Engineering Lab — mission scenarios (Apollo replay, interstellar route planning), live telemetry, and an AI console.
Most "AI for space" projects build a clever model and stop there. ARIA's hard part is the wiring — keeping an AI useful while making it incapable of doing something catastrophic on its own:
- A constitution above the AI. Forbidden actions, two-person rules, resource ceilings, and an Ed25519-signed sealed boot image live outside the model. The AI can propose; only the constitution can authorise. Nineteen explicit failsafes (F-1…F-19) cover prompt injection, token forgery, sensor spoofing, and the operator console itself.
- An independent monitor. A separate process — different code, different model family — votes 2-of-3 on every actuator-impacting command and can veto it or force safe mode. It never executes commands itself.
- One cognitive loop over 55+ tools spanning orbital mechanics, attitude control, anomaly detection, collision screening, life support, and radiation.
The five stages — propose → authorise → execute → observe → record — are split across process boundaries, so compromising the AI still leaves the constitution, the monitor, and a tamper-evident audit log intact, and any of them can force the craft into safe mode.
Fed reconstructed Apollo 13 cryo-tank telemetry, ARIA's full loop flagged the O₂ tank-2 pressure excursion 94 seconds before the historical master alarm — and the advisor returned an isolation plan matching documented EECOM procedure, including preparing the lunar-module "lifeboat" checklist ~13 minutes earlier than the real mission issued it. End to end in under 16 seconds. (Honest caveats + a reproducible run: docs/APOLLO13_REPLAY_REPORT.md.)
A research prototype — nothing here has flown. Read it as evidence of an architecture that could fly with the right partner and hardware, not something to launch tomorrow.
- ~1,300 Python modules across 35 subsystems; 488 test files; 19 coded failsafes
- React + WebSocket mission dashboard
- All bundled data is public-source (NASA, NOAA, CISA, ECSS) — no external IP or customer data
- Not there yet: flight heritage, DO-178C paperwork, a rad-hard target, an on-device model
System architecture · Failsafe design · Honest assessment · doc index · full original overview in docs/OVERVIEW.md
git clone https://github.com/dsremo/aria.git && cd aria
pip install -e .
aria --help # CLI dispatcher
# Optional: set ANTHROPIC_API_KEY for the LLM backend; otherwise a deterministic rule-based fallback runs.Source-available under the ARIA Source-Available License v1 — perpetual, with no change date and no auto-conversion to open source. Free to read, learn from, and use for personal/hobby and unfunded academic work; all commercial or production use requires a paid licence.
Built by Ashutosh Tiwari.