SO-101 Robot Arm: Build vs. Buy, and a Robotics Learning Path

Sources

Context and Motivation

Starting point for a robotics learning track, grounded in an undergraduate math background. First concrete decision: how to acquire the SO-101 arm (Hugging Face’s LeRobot reference platform) — source parts and 3D print from scratch, or buy a kit. No 3D printer access currently; priority is hands-on learning value from the build itself, not just fastest time-to-arm.

Platform Basics

The SO-101 system is a leader-follower teleoperation pair:

• Follower — the working arm. Most joints use STS3215 servos with 1/345 gearing (high torque, holds position under load).
• Leader — a puppet arm with deliberately weaker/mixed gearing (1/191, 1/345, 1/147 across the six joints — see table below) so it can be backdriven by hand with minimal resistance.

You physically move the leader; its joint encoder positions stream over USB (Feetech half-duplex TTL bus) as target commands to the follower. LeRobot’s imitation-learning workflow records (camera frames + joint state, action) pairs from this teleoperation, then trains a policy (ACT, diffusion policy, or a VLA) to reproduce the demonstrations.

Cost Comparison

Three viable paths, given no current 3D printer access:

Official BOM total (2 arms, US pricing, electronics only — no printed parts, no camera): $229.88, per TheRobotStudio’s BOM (12× STS3215 servos, 2× motor control boards, USB-C cables, power supplies, table clamps, screwdriver set). Single-follower-only equivalent: ~$122.

3D printing spec (if self-printing): PLA+, 0.4mm nozzle / 0.2mm layer height (or 0.6mm nozzle / 0.4mm layer), 15% infill. Verified on Prusa MINI+, Creality Ender 3, Bambu Lab systems. Minimum bed size ~220×220mm.

Vendor Comparison — Unassembled Kits (Path C)

Decision made: going with Path C, unassembled, so the arm is built and wired by hand rather than pre-assembled. Comparison below is scoped to unassembled/DIY-tier listings only — pair pricing, US buyer. Sorted by all-in cost (kit + camera if not bundled; ~$25 for a basic USB webcam where needed).

Read: Seeed Studio Pro is now the strongest option — same vendor sells both the motor kit and a dimensionally-matched printed-parts add-on, and it uses the original Feetech STS3215 servos, eliminating the cross-vendor compatibility risk that Hiwonder’s DIY kit introduces. Hiwonder and OZ Robotics both need their printed-parts claims independently re-verified (via a packing-list image or direct vendor contact) before being trusted at face value — this thread already caught one vendor-description error that a summarized page fetch missed, so treat unverified “includes printed parts” claims with real skepticism going forward.

Synthesis

Given the stated priority (hands-on learning over pure speed or pure cost) and no printer access: the dominant learning value in this build is in mechanical assembly, servo bus wiring, motor-ID/EEPROM configuration, and calibration — not in operating an FDM printer or researching individual BOM vendors. Those are real skills, but tangential to robotics specifically.

Decision: Path C, unassembled. Buy a complete unassembled kit (pre-printed parts + motors, ideally camera bundled) and do 100% of the assembly, wiring, motor configuration, and calibration by hand. This captures ~90% of the hands-on value of Path A at lower logistics overhead and comparable-or-lower cost, and defers the 3D-printer purchase decision to later — once it’s clear whether more hardware projects are coming, that’s a better-informed $200–250 call than making it upfront.

Revised vendor recommendation: Seeed Studio Pro (~$349 all-in). The Hiwonder recommendation above is retracted — its DIY kit is missing both the printed frame and a real camera, and pairing it with a third-party printed frame carries unverified fit risk given Hiwonder’s non-standard servos. Seeed sells a dimensionally-matched motor kit + printed-parts add-on from the same vendor, using the original Feetech STS3215 servos the whole SO-ARM100/LeRobot ecosystem is built around — no cross-vendor compatibility guesswork. Before ordering, still worth directly confirming Seeed’s own packing-list/product images the same way Hiwonder’s were checked here, rather than trusting the listing text alone.

Before ordering anything: verify a vendor’s “includes printed parts” / “includes camera” claims against an actual packing-list image or itemized spec sheet, not just page-summary text — that’s precisely how the Hiwonder error above was caught.

Open Questions

• Verify Seeed Studio’s own packing-list/product images directly (the way Hiwonder’s were checked) before trusting that its printed-parts add-on is confirmed complete and correctly matched to the motor kit.
• Confirm whether Seeed Studio’s “US Warehouse” checkout option changes shipping cost/time or just dispatch location, and whether the $220/$240 Standard/Pro pricing (CNX Software, presumably AliExpress) is actually purchasable anywhere reachable from the US at that price.
• If Hiwonder is reconsidered later (e.g. if it’s meaningfully cheaper once a matched frame is found): does a printed frame exist anywhere that’s actually dimensioned for HX-10HM/HX-30HM servos, and is the “compatible with LeRobot SO-ARM101” marketing claim for these servos verified at the protocol level (does pip install -e ".[feetech]" actually talk to them, or does Hiwonder require a separate SDK/fork)?
• OZ Robotics’ “3D-printed structural parts” claim is unverified against a packing-list image — same caveat as Hiwonder before the correction above.
• Live stock check needed on WowRobo’s Package 2 ($259, variant id 46588630630617) and PartaBot before ordering — both showed “sold out” at research time.
• Once the arm is built: what’s the first task/project to target for imitation learning (informs camera setup, gripper choice, workspace design)?