Skip to content

Buy instead of build?

Building is not the only option, and it is worth being clear about the trade-offs before you spend a weekend printing parts.

OptionCostNotes
This project~£20Printed parts, hobby stepper, magnetic encoder. Full source and docs.
AliExpress kits$100–$200Search results. Assembled mechanics, variable quality, usually no documented control stack.
Quanser QUBE Servo 2~£4,500Product page. Precision hardware, courseware, MATLAB/Simulink integration, support.
  • The whole stack is inspectable. Every constant, measurement and design decision is in the repository, including the ones that turned out to be wrong.
  • It is a real sim-to-real problem. The rig is imprecise enough that policies trained purely in simulation do not transfer, which is what makes the fine-tuning and distillation steps meaningful rather than ceremonial.
  • It runs untethered. The end state is a self-contained device with a learned controller on an 8-bit microcontroller.
  • Precision. A stepper driving a printed arm through a printed bearing seat is not a servo on a precision gearbox. Friction varies between rebuilds — which is why system identification is a required step rather than an optional one.
  • Repeatability between rigs. Your friction parameters will not match the ones in this repository, and the champion policy’s numbers were measured on one specific rig.
  • Support and courseware. There is no lab manual and no one to email.

The commercial rigs exist for good reasons: they survive a lab full of students, they come with material, and they behave identically across benches. If you need twelve rigs that all do the same thing on the same afternoon, buy them.

If you want one rig where a student can see and change every layer — the mechanics, the firmware, the reward function, the network — this is a better object to learn from, and it costs a rounding error by comparison.

Other people’s takes on the same problem, several of which informed this build: