Skip to content
IJZIsaac Johan Ziebarth

Team project · UQ Space · 2026

HAVOC-2 firmware and COSMOS ground station

about 540 m, every frame passed

The C++ flight firmware I wrote for UQ Space's HAVOC-2 flight computer, and COSMOS, the ground station I built, which the team runs at launches. On HAVOC-2's first flight, to about 540 m, every logged frame passed its cyclic redundancy check (CRC), and every intact radio frame matched the SD card byte for byte.

My role
Lead avionics engineer until November 2026. Developed HAVOC-2's C++ flight firmware (all 24 firmware commits) from June to November 2026, when the team moved to a consolidated flight computer. Built COSMOS; nobody else has committed to it
Team
UQ Space avionics team; teammates designed the HAVOC-2 board
When
HAVOC-2 firmware Jun – Nov 2026; COSMOS May 2026 – present
Tools
C++, Teensy 4.1, RFD900 radio, Python, React, WebSockets, OpenRocket for Monte Carlo ensembles
Code
Private: UQ Space team repositories
Live figurehavoc2-flight.html

The full write-up

  1. The problem
  2. What I did
  3. Result
  4. What went wrong
  5. Credits

The problem

The team needed a flight computer that logs a rocket's flight and streams it to the ground reliably, and a ground station that shows the flight live and keeps a record that can be checked afterwards.

What I did

HAVOC-2 firmware

  • The firmware read an IMU, a high-g accelerometer, a barometer, 2 magnetometers and GPS, logged the IMU at 50 Hz to SD, and streamed CRC-checked packets to the ground over an RFD900 radio.
  • After the low velocity estimate on the first flight, I compared 3 replacement designs in replay and chose a complementary filter that blends the barometer with the IMU.
  • To keep the flash stall out of flight, I changed the firmware to write flash only on the ground.
  • I tracked down the magnetometer short that had hidden the IMU.

COSMOS ground station

  • The radio, the flight computer on USB, a GPS tracker and the weather station each get their own Python process, which parses its stream and feeds a React page over WebSockets.
  • The page shows flight state, altitude, position and alarms, and speaks flight callouts. The same parser drives a terminal view for launch day.
  • Every session goes into a hash-chained log (each entry is tied to the one before, so any change shows) that can be verified and replayed afterwards.
  • Operator commands are rate limited, and the ones that change settings need confirmation. The tests run in CI on every push.
The same flight in 3D, at true scale and relative to the pad, drawn from the same logs and with the same rules: solid lines are measured and dashed lines modelled, with the climb in ink and the descent in orange. The faint line on the grid is its track over the ground. Drag, or use the arrow keys, to turn it, and replay it to follow the rocket at 4 times real speed.
The flight-phase detector in HAVOC-2's firmware as of October 2026, which I redesigned after the first flight and tuned on its logs and bench captures. It reports the phase in telemetry and fires nothing. |a| is the IMU's measured acceleration magnitude in g, relative to its own reading at rest, sampled at 50 Hz, and v is the vertical velocity from the filter that blends the barometer with the IMU.
How I put COSMOS together, as the code stands: one server process per device, one WebSocket per server for telemetry and commands, and a hash-chained session log that can be verified and replayed.
A small live replica of COSMOS's status pills, flight phases, altitude chart and event log, rebuilt for the site in COSMOS's own style. It replays COSMOS's built-in synthetic demo flight: made-up physics, not a real rocket. Pause it, change the speed or drag the slider.

Result

  • On its first flight, in July 2026, HAVOC-2 flew as the data recorder on a sub-scale rocket that reached about 540 m after an 11.7 g boost.
  • On the replayed flight log, the new estimator cut false speed on the pad from a peak of 137 m/s to 0.37 m/s, and with sensor faults injected it never called apogee early.
  • At the 2026 Australian Universities Rocket Competition, COSMOS recorded flight computer data from the launch rail and, about 3 hours before launch, ran a 512-run OpenRocket Monte Carlo ensemble from the weather station's pressure and wind to predict the apogee spread and the landing footprint.

What went wrong

  • Brownout: HAVOC-2 browned out at ejection on the first flight. Logging resumed, but about 9 s of the descent went unrecorded while it rebooted.
  • Low velocity estimate: on the first flight the onboard velocity estimate read about 5 times too low.
  • Flash stall: erasing the onboard flash stalled the clock for 225 ms every 5 s.
  • Magnetometer short: a shorted magnetometer breakout took down the I2C bus and hid the IMU behind it.
  • No flight data yet: the new estimator sat on the launch rail at the 2026 competition, but the rocket broke up 2.8 s into the flight and the nosecone holding the recorders has not been found.

Credits

HAVOC-2's board was designed by teammates in the UQ Space avionics team; the page covers its firmware and the ground station.