The number that jumps out is the error drop. On a 45:1 3D-printed reduction gearbox, the settled joint error went from 0.065 degrees down to 0.014 degrees, a 78% reduction, just by feeding back the position measured after the gearbox instead of trusting the motor alone. For a board the maker wants to sell for around $20, that result is remarkable.
The board, called CLMD, is built around an ESP32-S3 and a Trinamic TMC2209 driver. Two TLE5012B magnetic sensors handle the dual encoder inputs, one reading the motor shaft angle and the other reading the gearbox output angle. Everything was drawn up in EasyEDA, the PCBs were ordered through PCBWay, and firmware flashes over USB-C using a desktop tool called Studio. The finished board is blue, ringed with white multi-pin connectors, with a USB-C port on one edge and two push buttons.
Plenty of small choices stand out. There is reverse-polarity protection and native USB flashing, and the connector set covers two CAN bus plugs, two endstops, VIN, 5V, motor and GPIO. In testing, CAN was daisy chained across six arm joints, and the maker says hooking up six boards should be easy thanks to that layout. It works with NEMA 17 and NEMA 23 motors. One catch on cost: while the target is $20 per board, small batches of 5-10 came to almost $40 each.
The clearest trade-off is the driver itself. The TMC2209 cannot do FOC, which the maker calls the biggest downside, and 48V operation is off the table with it. There is also an open question the maker raised: handling backlash compensation without destabilising the control loop.
Project by Arctos Robotics. Watch it on their YouTube channel: