SO-101 Manipulator

A self-directed project teaching myself robotics on a low-cost open-source arm (SO-101 / HuggingFace LeRobot). The rule I set myself: write every layer from scratch rather than driving the arm through a stock pipeline. The goal is to understand how a manipulator actually works, not to run someone else’s demo.

What’s built

Servo tooling. Bus scanning, ID reassignment, live joint readout. The first real bug was a duplicate servo ID — two motors answering to the same address — which is why the arm wouldn’t calibrate at all.

Teach-and-repeat. Torque off, hand-move the arm, record all six joints at 50 Hz to JSON; torque on, replay at the recorded timing.

An Arm class. One reusable hardware interface — port lifecycle, torque, read/write, tick↔radian conversion, and smooth interpolated motion. Every earlier script had reimplemented this; now the endianness fix lives in exactly one place.

Calibration. Each joint’s zero derived from its measured mechanical range, cross-checked against the URDF’s declared range so a bad capture is caught at the arm instead of three steps later.

Forward kinematics. A chain of homogeneous transforms from joint angles to end-effector pose, built on CAD-derived URDF parameters.

Result: FK validated against measurement

Predicted position checked against ruler measurements at four poses spanning 14–40 cm of reach:

Mean absolute error	0.8 cm
Systematic bias	+0.27 cm in x, +0.66 cm in z
Residual scatter after removing bias	0.37 cm (x), 0.44 cm (z) RMS
Correlation of error with reach	−0.94

The constant bias in z turned out to be a datum-placement error — the measurement reference sat about 7 mm below the URDF’s base plane — not a fault in the model. The row that matters is the last one. A wrong joint zero is an angular error, so it gets multiplied by the lever arm and grows as the arm extends. Here the error shrinks with reach: the two most extended poses were the two most accurate. That rules out joint-zero error as a contributor.

Residual scatter of ~0.4 cm is consistent with what hand measurement can resolve — locating the wrist-roll axis by eye is itself worth ±0.3–0.5 cm.

Things that cost real debugging time

• The servos lie about byte order. Raw 16-bit reads come back byte-swapped, and writes need the same correction. Straightforward once found; baffling until then.
• /dev/ttyACM* numbering is not stable. It reshuffles with USB enumeration order, so “the wrong arm is moving” is a real failure mode. Both arms are now pinned by USB-adapter serial.
• Replay smoothness is about distance per hop, not hop count. Interpolation steps that are too fine make the servo finish each hop instantly and then idle until the next command — which reads as jitter. Deriving the step count from travel distance makes small and large moves glide alike.
• Calibration has to agree with the URDF’s convention. The URDF defines joint zero as the middle of each joint’s range, not a straight pose. A hand-posed “straight home” reference was off by ~12° on two joints, and the kinematics were correct while disagreeing with reality the whole time.

Next

Servo characterization (accuracy, latency, sag under load), a hand-written PID joint controller, then a camera rig and a learned visuomotor policy trained on teleoperated demonstrations.