From the vault’s bluntest [reference] label: this material validates against an experiment, not a compiler. These projects are the cheapest available substitute — simulate it, then check the simulation against something physical you can measure.

Difficulty: 🟢 starter (hours–days) · 🟡 intermediate (a week or two) · 🔴 ambitious / portfolio-grade. ⭐ = highest signal.

Python + NumPy/SciPy + Matplotlib is the whole toolchain. Nothing here needs licensed software.

Projects

  • 🟢 ⭐ Tune a PID on a simulated plant — a first-order-plus-dead-time model, then tune by Ziegler–Nichols and by hand. Plot step responses for each gain change. Done when: you can produce overshoot, sluggishness and instability on demand and explain which term caused each. Exercises: PID.

  • 🟢 Draw the plots yourself — root locus, Bode and Nyquist for the same system, in code. Done when: you can predict from the Bode plot what the step response will look like, then confirm it. Exercises: Bode.

  • 🟡 ⭐ Balance a real thing — a physical inverted pendulum or a reaction wheel with a cheap microcontroller and an IMU. Done when: it balances, and your measured settling time is within sight of your simulation’s. This is the rep that closes the gap this folder names — theory meeting a real actuator with real noise and real latency. Ties to hardware and robotics.

  • 🟡 A Kalman filter on noisy real data — log accelerometer data from a phone, fuse it, and compare against a complementary filter. Done when: you can show the filter tracking through a period where the raw signal is useless. Exercises: observers.

  • 🔴 MPC on a constrained system — model predictive control with actuator limits, against the same plant you PID-tuned. Done when: you can show MPC respecting a constraint that PID violates.

If you only do one

Balance a physical inverted pendulum. It is the only project here that validates against reality rather than against another equation.

↑ Back to Control Theory