Control

The oldest controller in the book and still the most used. How it works, how to tune it, and when it's the wrong tool.

~15 min

PID works for one variable. State-space and LQR are the principled way to control many variables at once, with one cost function instead of six tuning knobs.

~17 min

Every control tick, solve an optimization over the next N timesteps, apply the first action, repeat. The controller that took over autonomous driving, humanoid gait, and drone racing.

~18 min

Most robots are nonlinear. Most controllers prefer linear. Feedback linearization is the trick that turns nonlinear robots into linear ones — for free, on the inside — so you can apply LQR or PID on top.

~14 min

The moment a robot touches the world, position control breaks. Impedance control turns the arm into a programmable spring; admittance control inverts the relationship. Here's the difference, the math, and when to pick which.

~14 min

Impedance lets the arm be soft. Force control makes it push with exactly the right force. Hybrid combines them — position-controlled on some axes, force-controlled on others. The architecture behind real industrial assembly.

~14 min

Close the control loop on the camera, not on joints. Image-based servoing controls in pixel space; position-based servoing controls in 3D pose. Same goal, different math, different failure modes.

~13 min

The energy-function trick that turns 'I think this controller works' into 'I can prove it converges.' Useful exactly when intuition runs out — nonlinear systems, adaptive control, sliding-mode.

~13 min

Acrobots, cart-poles, walking robots — all have more degrees of freedom than actuators. You can't dictate every joint. Instead you exploit the dynamics. Energy shaping is the trick that does it.

~14 min

1 kHz control loops, hard deadlines, deterministic timing. The kernel tricks, scheduler tweaks, and architecture choices that separate hobby control from production.

~14 min