Robots you can build yourself

The hardware ladder

The SO-ARM100 and its successor the SO-101 are open-hardware arms: the parts are printed, the motor kits are bought, and the assembly takes a few hours. LeRobot, the Hugging Face robotics stack, treats them as first-class, so recordings drop straight into an imitation-learning pipeline.

Prices below were gathered for the survey on 11 August 2026 and are a snapshot, not an offer. Check the seller before ordering, and note that printed parts and tariffs are not included in most of them.

• Established

  SO-101 motor kit, standard, $220. Pro kit $240. Complete kit around $269. Printed parts roughly $35 more.

• Established

  LeKiwi, a three-wheeled base with a Raspberry Pi inside, from $260. It turns the arm into a mobile manipulator.

• Established

  SO-101 is 6-DOF and takes roughly three to four hours from parts to first data collection, print time included.

• Established

  LeRobot supported robots as of June 2026: SO-100, SO-101, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy 2, Unitree G1, OpenARM, reBot B601.

• Lead, unverified

  ROS 2, Nav2, MoveIt 2, micro-ROS and Dora-rs were not retrieved for the survey. No version claims are made about them here.

SO-ARM100 (The Robot Studio) LeRobot: SO-101 docs LeRobot: assembling the SO-101 LeRobot: SO-100 docs SO-101 kit (Seeed Studio) SO-ARM100 kit (Seeed Studio)

Hands, and the one costed here

An arm needs a hand before it can do useful work, and the hand is the harder part to build. The ORCA Hand is the open design studied most closely here: seventeen degrees of freedom, tendon driven, with an official bill of materials and printable parts.

A full build pack for it was worked out in September 2026 against the printer in this workshop, a Bambu Lab X1C with a 0.4mm nozzle, for a right hand with no tactile sensing. Everything below comes from that pack. Prices are source snapshots from that date, not checkout totals, and the sellers should be checked before anyone orders.

• Established

  Two motor routes, and the difference is most of the cost. Dynamixel: sixteen XC330-T288-T finger servos at $103.39 each plus an XC430-T240BB-T wrist servo at $137.89, a PowerBoard at $21.85 and a U2D2 converter at $36.92. Feetech: sixteen HL-2915-Plastic finger servos plus one HL-3930-Plastic wrist servo at about $76, with an FE-URT-2 USB-C control board at about $19.99.

• Established

  The rest is small parts and quantities that are easy to get wrong: 59 bearings at 4x8x3, two at 6x13x5, 16 magnets at 3x5, a 142mm GT2 belt, roughly 70 M2x4 screws, and 30m of UHMWPE fishing line at 0.39mm for the tendons.

• Established

  Materials: PLA+ for the rigid parts, TPU around 85A for one mould plate, and Dragon Skin 10 NV for the silicone skins, which are cast rather than printed.

• Lead, unverified

  The exact Feetech part is HL-2915-C001 in the plastic housing. The pack flags this as needing confirmation and says not to substitute the STS3215.

• Sources conflict

  The Lite model has files but its 3MF is still listed as to be decided, so Lite quantities must not be mixed into a v2 order. Read the two as separate builds.

ORCA Hand: bill of materials Hardware, STL and 3MF files orca_core, the controller (MIT) Assembly video ORCA Hand Lite AmazingHand (Pollen Robotics) qubit (0xaiwhisperer)

One warning

From the build pack, unedited: reversing the polarity on a Feetech connector can destroy the servo. Never force one.

With sixteen finger servos in one hand, this is the most expensive mistake the build allows, and it takes a second to make.

The software is already open

orca_core is MIT licensed and does the hardware control, motor configuration, tendon tensioning, calibration and the joint-space API. Everything else is an optional layer on top: robot descriptions for simulation, a simulation environment, teleoperation and retargeting, and a web view for tactile sensors on the models that have them.

Setup is a git clone, uv sync, adding yourself to the dialout group, and running the motor chain configuration script with the Feetech flag.

orca_core

Controlling a hand with your own muscles

MIT published a wristband that reads muscle activity with ultrasound and drives a robotic hand in real time. It shows one way to steer a robot hand without holding a controller. The paper is free to read.

MIT News: wristband controls a robotic hand The paper (ACM) Summary (humanoid.guide) MIT News: xStrings, printing things that move

What the benchmarks do not tell you

The survey concludes that hardware no longer limits home robots. What limits them is the long tail of situations a model has not seen.

• Established

  Payload describes what a robot can do better than its degrees of freedom. Battery life bounds everything: a Unitree G1 runs around two hours mixed-use on a 1.5 hour charge.

• Established

  Home manipulation today depends on teleoperation. The only home humanoid on sale is driven by a person some of the time.

• Established

  Knowing when to stop, the abstention problem, is called out in the survey as the most important result in its foundation-model section.

• Commonly claimed, not established

  Benchmark numbers taken at face value. The survey has a section on how far they can be trusted and concludes that they overstate real ability.