RTOS and embedded AI

Why this matters

A flight controller must do many jobs at once: read the IMU a thousand times a second, compute attitude, receive GPS, write logs and talk to the ground station. If logging runs so long that an IMU read misses its time, control degrades immediately. A real-time operating system (RTOS) orders the work so important tasks always run on time. The final topic leads on to running small AI models on MCUs.

Tasks, priority and preemption

• A task is work repeated periodically, with period $T$ and computation time $C$
• Priority is the task’s importance
• Preemption lets a more important task interrupt a running one at once

Rate-monotonic scheduling gives higher priority to shorter periods. Liu and Layland (1973) proved that if total utilisation $U$ does not exceed $n(2^{1/n}-1)$ for $n$ tasks, every task meets its deadline. This bound falls towards about 69% for many tasks. It is a sufficient condition: exceeding it does not necessarily mean failure, and needs a closer check.

def check(tasks, horizon=10_000):
    n = len(tasks)
    u = sum(c / t for _, t, c in tasks)
    bound = n * (2 ** (1 / n) - 1)
    order = sorted(tasks, key=lambda x: x[1])            # shorter period = higher priority
    remaining = {name: 0 for name, _, _ in order}
    missed = {name: 0 for name, _, _ in order}
    for t in range(horizon):
        for name, period, cost in order:
            if t % period == 0:
                if remaining[name] > 0:
                    missed[name] += 1                    # previous job unfinished = missed deadline
                remaining[name] = cost
        for name, _, _ in order:
            if remaining[name] > 0:
                remaining[name] -= 1
                break
    return u, bound, missed


base = [("IMU", 10, 2), ("attitude", 25, 5), ("logging", 200, 40), ("GPS", 1000, 50)]
for label, tasks in (("base", base), ("+ AI 6 ms/20 ms", base + [("AI", 2000, 600)]),
                     ("+ AI 8 ms/20 ms", base + [("AI", 2000, 800)])):
    u, bound, missed = check(tasks)
    print(f"{label:<16} U = {u:.2f}, bound = {bound:.3f}, missed deadlines = {sum(missed.values())}")

base             U = 0.65, bound = 0.757, missed deadlines = 0
+ AI 6 ms/20 ms  U = 0.95, bound = 0.743, missed deadlines = 0
+ AI 8 ms/20 ms  U = 1.05, bound = 0.743, missed deadlines = 4

RTOSs in flight stacks

• PX4 uses NuttX as its primary RTOS on flight-control boards. NuttX is an Apache project under the Apache 2.0 licence
• ArduPilot on STM32 boards uses ChibiOS through the AP_HAL_ChibiOS layer. ChibiOS/RT is GPL3 or commercial, while ChibiOS/HAL is Apache 2.0
• FreeRTOS is popular on general MCU boards, including the ESP32; its kernel is MIT-licensed

Besides scheduling, a safe embedded system needs a watchdog, which resets the system if the program hangs and fails to check in on time.

From MCU to AI

TinyML runs small machine-learning models on MCUs, for example classifying abnormal motor sounds or vibration patterns. Google calls its tool LiteRT for Microcontrollers (formerly TensorFlow Lite for Microcontrollers); it runs without an operating system, and its core runtime fits in about 16 KB on a Cortex-M3. Larger tasks, such as detecting objects in images, need a companion computer as in UAT 322, and the knowledge hub’s deep dive on embedded AI and ROS builds on this module.

Module lab

Common mistakes

Summary

• An RTOS orders tasks with priority and preemption so important work meets its deadlines
• Rate-monotonic gives shorter periods higher priority; the $n(2^{1/n}-1)$ bound is a sufficient condition
• PX4 uses NuttX, ArduPilot uses ChibiOS, and FreeRTOS is popular on general MCUs
• TinyML runs small models on MCUs, while larger AI needs a companion computer

Check your understanding

1. A task with a 10 ms period takes 2 ms and one with a 20 ms period takes 5 ms. What is the total utilisation?
2. What is the rate-monotonic bound for 2 tasks?
3. Is the task set in question 1 guaranteed to meet its deadlines?
4. Which RTOS does PX4 use on flight-control boards?
5. What is a watchdog for?