Imagine a machine that does for robots what a 3D printer does for plastic parts: you send it a design, and a working robot comes out. If it existed, people could get useful machines for little more than the cost of materials. This post asks what such a machine would have to be, starting from one measured fact about today's cheapest open robot arm.
Nothing here was built or tested on hardware. It is a direction for research, not engineering or business advice.
The plastic is the cheap part
The SO-101 is an open-source robot arm from TheRobotStudio and Hugging Face, popular for robot-learning experiments. Its frame is 3D printed and everything else is bought. Its repository lists the bought parts for one arm at $121.94 (US prices) and publishes the printed parts as STL files.
We downloaded the follower arm's STL and measured it with a short script (so101_cost.py). The printed parts enclose 571 cm³. Solid PLA would weigh 708 g. At the repository's recommended 15% infill, the estimate is 277 to 447 g, depending on wall thickness. That is a geometric estimate; we did not run a slicer.
| Line | Cost | Share |
|---|---|---|
| Six STS3215 servos | $83.34 | 70.5% |
| Controller board, USB cable, power supply | $27.60 | 23.4% |
| Printed plastic (about 360 g at $20/kg) | about $7.24 | 6.1% |
Tools such as clamps and a screwdriver are left out. Even if every part were printed solid in $30/kg filament, the plastic would come to $21. The servos would still cost four times as much.
So a machine that "prints robots" faces little difficulty with the printing. The structure of a hobby robot is already cheap. What costs money is the motion: motors, gears, position sensors and the electronics that drive them.
Can we print the motors too?
Researchers are working on it, and the results are impressive but early.
- In 2023, a team from ETH Zurich, MIT and the company Inkbit printed a robotic hand with bones, ligaments and tendons in a single job, using an inkjet printer that scans each layer with cameras (Buchner et al., Nature, 2023). The hand is a mechanism; it still needs something outside to pull the tendons.
- In 2026, MIT researchers reported a fully 3D-printed electric motor: a linear actuator made from five materials in about three hours, with about 50 cents of material (MIT News). It moves 318 micrometres at its resonant frequency, and its magnets must be magnetised after printing. That is a real first, but it is a long way from the 16.5 kg·cm of stall torque a $14 hobby servo provides.
Fully printed robots will come. For now, a system that makes working robots cheaply has to buy its actuators.
What the factory would actually look like
If printing is cheap and actuators must be bought, the "robot printer" becomes a small automated cell with five parts:
- A farm of ordinary 3D printers for the structure. This part is solved.
- A short catalogue of mass-produced modules that every design uses: smart servos that chain on a single cable, one controller board, one power supply and standard screws. Buying a few part types in volume is where the low price comes from.
- A robot arm that assembles the robot: it seats servos in printed pockets, drives screws and plugs in cables. Daisy-chained servos matter here because routing loose wires is one of the hardest jobs to automate. Designs would be rewritten so that everything goes in from above, with snap fits instead of screws where possible.
- A self-test in which the finished robot calibrates and checks itself with its own motors before it is packed.
- Later, more printed parts: printed gearboxes around bought motors, then printed motors, each adopted when it beats the catalogue part on cost and strength.
For circuit boards, open-source desktop pick-and-place machines such as LumenPnP can already place small surface-mount parts. The hard step in between is the one people do by hand today: fitting modules together and wiring them.
What we already know from simulation
This repository's research portfolio has studied pieces of that cell in simulation. The robot assembly roundup covers them. Two results bear directly on the design:
- In a robot-plus-printers cell, the arm that loads and unloads the printers limited throughput more than the printers' utilisation suggested.
- An arm that can only work from above could still assemble a toy gripper, in 120 of 720 possible part orders, provided the design allowed it.
The self-replication roundup found the same pattern in a RepRap 3D printer: motors and electronics dominate what a shop cannot make for itself.
What comes next
The next studies use the real arm instead of toy shapes: count the SO-101's real assembly steps and which of them a robot could do, simulate a cell that builds SO-101s to estimate arms per day and labour per arm, and price a printed gearbox against the $13.89 servo. The plan is in the agent brief.