14 notes, ~23,000 words, and the README is blunt that nothing here has been validated on hardware. These projects exist to change that — and simulation counts, because Gazebo has real dynamics even if it has no real friction.

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

The ladder

  • 🟢 Make one motor go exactly where you tell it — a servo or a stepper with an encoder, commanded to a position and holding it. Then disturb it by hand and watch it correct. That’s a closed loop, and it’s the entire subject in miniature.
  • 🟡 ⭐ Tune a PID by hand, and write down what each term did — a line follower or a self-balancing two-wheeler. Start with P only and watch it oscillate; add D and watch it stop; add I and watch the steady-state error close. Tuning one loop badly on real hardware teaches more than any amount of control theory reading — and it’s the honest prerequisite for trusting a word of the PID note.
  • 🟡 Sensor fusion on something that moves — combine an accelerometer and a gyro into one angle estimate with a complementary filter, and demonstrate why neither alone is usable (one drifts, one is noisy). The intuition Kalman filters formalise.
  • 🔴 Teleoperation with a latency budget — drive something remotely and measure the end-to-end delay, then make it degrade safely when the link drops rather than continuing at the last command. Joins latency to a machine that can hurt someone.
  • 🔴 A robot that maps a room — odometry plus a range sensor into a 2D occupancy map, and an honest account of how far it drifts. The entry point to SLAM, and the project that turns “odometry drifts without bound” from a sentence you read into a number you measured.

If you do one: the PID tuning project. It’s cheap, it fits on a desk, and it’s the difference between having read the PID note and having earned it.

If you only do one

Get something moving in ROS 2 + Gazebo. Simulation removes the hardware cost while keeping the coordinate frames, the timing and the tuning — which is where the actual learning is.

↑ Back to Robotics