Shreyas

UMBRA is a lunar exploration simulator built to test a harder question than whether AI can write code: can it understand human intent, act independently, and recognise when to stop? Two GPT-6 Astra agents coordinate as Mission Control and an onboard rover across a simulated Earth–Moon communication delay. A commander gives a plain-English goal—“Find water ice near the shadowed basin without risking the drill”—and approves the mission. Astra chooses where to explore, adapts when terrain becomes hazardous, and refuses missions it cannot safely return from. Built on real lunar elevation and imagery distributed through NASA’s Moon Trek, UMBRA combines calculated illumination with simulated ice deposits and rover operations. The model receives terrain and illumination data, but never the hidden ice map: it must reason about where to search. I built UMBRA to explore what trustworthy autonomy looks like where human intervention cannot arrive immediately. The defining demonstration is not just a rover completing a mission—it is a rover explaining why an unsafe mission should not begin.