Build an autonomous greenhouse
A closed loop that runs for six weeks unattended, in a system where failure is measured in dead plants.
What this is
A Launch capstone with a property none of the others have: the feedback is biological, slow, and unforgiving. You cannot roll back a dead tomato plant.
It is also the most honest test of unattended operation in the catalogue. A system that works while you watch it is not the deliverable — a system that works for six weeks while you do not is.
What you need
- Somewhere enclosed and growing. A greenhouse, a polytunnel, a grow tent, or a sunny windowsill with four pots. Scale down freely; the loop is the same.
- A microcontroller, a soil moisture sensor, a temperature and humidity sensor, and one pump or valve. Around £40 all in.
- Plants you would be sad to lose. This matters — it is what makes you design the failure modes properly.
What you'll leave with
- A fortnight of baseline environment data before anything actuates.
- A closed loop with bounded actuation that fails safe when a sensor lies.
- Six weeks unattended, with the intervention count to prove it.
- An outdoor scale-out plan where you have control of far less.
The loop
Sense, then decide, then actuate — in that order, over about ten weeks.
- Two weeks of sensing nothing but data — Log before you act. The setpoint you would have guessed is almost certainly wrong. (30 min)
- The loop that cannot drown them — Closing the loop is easy. Bounding it so a lying sensor cannot destroy anything is the actual work. (35 min, paid)
- Six weeks away, then outdoors — The unattended run that proves it, and the scale-out where you control almost nothing. (35 min, paid)