3D Printing an SO-100 Arm That Holds Tolerance

The SO-101 repository ships a small part whose only job is to tell you whether your printer is good enough to build the arm. Slice Gauge_0.STL halfway up and the pocket inside measures 34.900 mm by 24.700 mm. The Feetech STS3215 servo that has to drop into it is 24.7 mm wide. There is no designed clearance on that axis. If your walls land 0.05 mm proud of nominal, the servo does not go in.

That measurement is the whole story of printing an SO-100 or SO-101. The geometry is not hard. What the design demands is that a 25 mm feature comes out within about a tenth of a millimetre, roughly 0.4 percent. Most printers are not there out of the box, and the failure does not look like a print failure. It looks like an arm you cannot assemble, or an arm you can assemble that has play in the wrist and records noisy episodes.

The fits that actually decide anything

I pulled the current STL files from the build repository and measured the meshes directly rather than trusting the README. The method is in the callout below.

The M3 row is the one that eats an evening. A 3.20 mm hole for a 3.00 mm screw leaves 0.10 mm of air per side. A printer laying its walls 0.10 mm proud on each side of an internal feature, entirely normal for an uncalibrated machine, takes that hole down to 3.00 mm, which is the screw's own major diameter, and it will not pass. Forcing it splits the boss along a layer line, and you will not see the crack until the end effector starts wobbling weeks later.

The published settings, and which document you are reading

Three sets of recommended settings live in the same repository and they do not agree. Here they are side by side.

That is not a contradiction so much as three risk appetites. 13 percent was the original SO-100 figure, 15 percent is current, and print bureaus quote 20 percent because they would rather over-build than field a complaint. None of the three is why your parts do or do not fit. Infill density has almost nothing to do with dimensional accuracy at this size.

If you are building the SO-100 you are following a document the maintainers marked deprecated. Everything here applies to both, and the SO-100 against SO-101 comparison covers what changed. The visible difference in the print files is the leader, which now uses three different gear ratios rather than one.

git clone --depth 1 https://github.com/TheRobotStudio/SO-ARM100.git
cd SO-ARM100

# the gauges you print first
ls STL/Gauges/
# Gauge_0.STL  Gauge_tight_1.STL
# Lego_Size_Test_02_zero.STL  Lego_Size_Test_02_minuspoint1.STL

# pre-arranged plates, already oriented Z-up to minimise supports
ls STL/SO101/Follower/    # Ender_Follower_SO101.stl  Prusa_Follower_SO101.stl
ls STL/SO101/Leader/      # Ender_Leader_SO101.stl    Prusa_Leader_SO101.stl

# individual parts, for reprinting one at a time
ls STL/SO101/Individual/

The plates use nearly the entire bed

The combined plates are convenient and they are also a trap. Measured from the mesh bounding boxes:

There is a second reason to split the plate. Both are 87 to 88 mm tall, so a single failure eight hours in wastes everything on it. Batches give you a restart point, and let you check the first parts against a real servo before committing the rest of the roll.

Material: PLA, PLA+ and PETG, with real numbers

Upstream says PLA+, which is good advice with the reasoning left out. PLA+ is not a standard, it is a vendor blend, and every vendor means something different by it. Prusa's own trade name for the material below is Prusament PLA/PLA (Blend). What matters is the property set. Both columns come from the Prusa Polymers datasheets, version 1.1 dated 16-02-2022. The mechanical rows are measured on printed specimens; hardness and interlayer adhesion come from the filament table on the page before.

Tensile modulus decides it. PLA is 2.3 GPa, PETG 1.5 GPa, a 35 percent stiffness drop. On a link carrying the degrees of freedom above it, that shows up as sag under load, and sag is a systematic error baked into every LeRobot dataset you record. The servo reports its commanded angle, not where the gripper actually is, so a flexing forearm is invisible in the data and very visible in the results.

Layer height, walls and infill: which knob does what

These three get discussed as if they trade off against each other. They mostly do not. Only one of them has much to do with whether parts fit.

If you take one thing from this section: 4 walls with 10 percent infill is stiffer in bending than 2 walls with 40 percent, and uses less filament. Bending stiffness scales with how far material sits from the neutral axis, and walls are the material furthest out. Use 3 walls on the links, 4 on the motor holders and base, and leave infill at 15 percent. At a typical 0.45 mm extrusion width that is 1.35 mm and 1.80 mm of solid shell.

Raising wall count does cost you something on tolerance. Every extra perimeter is another chance for the slicer to over-pack plastic into an inside corner, and the servo pockets are full of inside corners. That is what the compensation settings are for, and it is why you print the gauges after choosing your wall count, not before.

Print the gauges before you print the arm

This is the step almost everyone skips and it is the highest-value fifteen minutes in the build. Four gauge files ship in STL/Gauges/: two check against a real STS3215, two against a standard 4x2 Lego brick for people whose servos have not arrived.

1. 1
   Slice both servo gauges with your real profile

   Same material, nozzle, layer height and wall count you will use for the arm. A gauge printed with a different profile tells you nothing.

   bash

   STL/Gauges/Gauge_0.STL          # 34.900 x 24.700 mm pocket
   STL/Gauges/Gauge_tight_1.STL    # 0.100 mm tighter on both axes
   
   # if your servos have not arrived yet
   STL/Gauges/Lego_Size_Test_02_zero.STL
   STL/Gauges/Lego_Size_Test_02_minuspoint1.STL

2. 2
   Print both on the same plate in one run

   Printing them separately adds run-to-run variation and defeats the point. Together they take well under half an hour.

3. 3
   Read the result the way it was designed to be read

   Gauge_0 should accept the servo with firm hand pressure and no tool. Gauge_tight_1 should not. Both fit means you print undersize, neither fits means oversize, and only Gauge_0 fitting means you are inside 0.1 mm and can print the arm.

4. 4
   If both fit, add material

   Increase XY compensation in the positive direction in 0.05 mm steps and reprint Gauge_tight_1 only.

   ini

   ; PrusaSlicer, Print Settings > Advanced
   xy_size_compensation = 0.05
   
   ; OrcaSlicer / Bambu Studio, Quality > Precision
   xy_contour_compensation = 0.05
   xy_hole_compensation = 0.05

5. 5
   If neither fits, remove material

   The more common direction with a fresh profile. Go negative in the same steps and check flow rate before you pass -0.10 mm. Compensation trims hundredths, it does not hide a 5 percent over-extrusion problem.

   ini

   ; PrusaSlicer
   xy_size_compensation = -0.05
   elefant_foot_compensation = 0.2   ; first layer, 0.4 mm nozzle
   
   ; Cura, Shell settings
   xy_offset = -0.05                 ; "Horizontal Expansion"
   xy_offset_layer_0 = -0.2          ; "Initial Layer Horizontal Expansion"
   hole_xy_offset = 0.05             ; "Hole Horizontal Expansion", note the sign

6. 6
   Only now slice the arm

   Carry the same profile across. If you change filament brand or colour later, reprint Gauge_tight_1 first: pigment loading changes flow more than people expect.

Compensation settings, by slicer, with the real keys

The names differ in every slicer and one pair has an inverted sign convention, which is a reliable way to spend an afternoon going the wrong way. These come straight out of each project's current source, fetched 2026-08-24.

One myth worth killing. People blame small holes on the STL approximating circles with flat segments, and the geometry is real: a hole drawn with n segments is only d times cos(pi/n) across the flats. But the SO-101 screw holes are drawn with 24 segments, one vertex every 15.000 degrees.

So if your M3 holes come out 0.15 mm tight, the faceting in the file accounts for about 0.027 mm of that, under a fifth, and your extruder accounts for the rest. Fix flow and compensation rather than hunting a higher-resolution export. OrcaSlicer's hole_to_polyhole, labelled "Convert holes to polyholes" and off by default, runs the other way: it replaces an almost-circular hole with a deliberately low-vertex polygon, scaled up by 1 / cos(pi/n) so the flats sit tangent to the intended circle. That is the HydraRaptor polyhole trick, and it targets holes that come out small because of extrusion, not faceting that is already in the mesh.

Why the gripper is where tolerance turns into behaviour

Every joint has some print-related slop, but only the gripper converts it directly into task failure. It sits at the end of the chain, so every upstream error adds to it, and it is the only joint whose job is a binary outcome. A shoulder half a degree off still reaches the object. A gripper half a degree short of contact drops it.

The gripper claw is secured with 4 M3x6 mm screws on both sides, and the smallest bores in these parts measure 1.50 mm, which is where the M2x6 mm screws cut their own thread. Both are places where a bad print produces play rather than an obvious failure.

• Jaw screws in oversized holes: the jaw rocks under load, actual aperture differs from the reported angle by millimetres, and a policy trained on that data learns a mapping that stops being true once the screws settle.
• M2 bosses stripped by forcing screws into tight holes: the servo shifts under lateral load, which is one mechanism behind a joint that stops early.
• Servo pocket a fraction loose: the motor rotates a degree or two inside its mount on every direction reversal, your calibration drifts between sessions, and you will blame the software.
• The optional TPU 95A jaw from the June 2025 LeRobot hackathon forgives contact error, but it deforms under load, so grip force stops being a clean function of joint angle.

If you are chasing a gripper that does not close or an arm that twitches then sags, eliminate the mechanics before touching the model. Hold the arm in a fixed pose, command a small angle change on one joint, and watch whether the joint moves before the reported angle does. Backlash you can see is backlash that is in your training data. The failure-mode index lists which symptoms have mechanical roots.

Two routes to a working arm

pip install -e ".[feetech]"

lerobot-find-port

lerobot-setup-motors \
    --robot.type=so101_follower \
    --robot.port=/dev/ttyACM0

lerobot-calibrate \
    --robot.type=so101_follower \
    --robot.port=/dev/ttyACM0 \
    --robot.id=my_follower_arm

What a better printer does not fix

Dimensional accuracy is not repeatability. You can print an arm that assembles perfectly and still have a degree or two of backlash at the wrist, because backlash lives in the servo gearbox and the horn spline, not in the printed part. The STS3215 is a 55 g hobby-grade bus servo with a metal gear train. It is remarkably good for the money. It is not a harmonic drive, and no slicer setting changes that.

Reprinting a part mid-project quietly invalidates your data. Record two hundred episodes, crack a motor holder, reprint it with a different compensation value, and the relation between joint angles and end effector position has changed. Old and new episodes now describe two subtly different robots, and the model averages them. If you must reprint a structural part, note the date and treat what came before as a separate dataset. The same discipline runs through the guide to collecting high-quality VLA training data.

If you would rather generate data than filament dust, the operator route is the version where somebody else owns the mechanical problem. The complete SO-100 guide walks the same path from parts to a trained policy.

Questions people actually ask