Open-source humanoid robots you can actually build, with real parts costs

Yes, you can build an open-source humanoid robot that walks, from published files and parts you order online. Several teams made that possible on purpose, and their bills of materials are public. The surprise is the arithmetic: the cheapest well-documented printable humanoid that walks costs about what the cheapest humanoid you can simply buy costs. A 3D printer removes the machine shop from the problem. It does nothing about the motors.

The 3D-printed humanoid projects, side by side

Project	Size	Joints	Published parts cost	What it does
Berkeley Humanoid Lite	0.8 m, 16 kg	22 actuated	$4,312 US, $3,236 China	Walks, trained controller included
ToddlerBot (Stanford)	0.56 m, 3.4 kg	30 active	under $6,000	Walks, manipulates, teleop interface
Poppy Humanoid	83 cm	25 actuators	$8,000 to $9,000	Research and education platform
InMoov	life size	around 30 servos	not published by the project	Head, torso, arms. Legs not motorized
Zeroth Bot	not stated on the repo	not stated	from $350	Public alpha, sim-to-real research toy
Open Duck Mini v2	around 42 cm	not stated	under $400	Bipedal, no arms. Not a humanoid

Where the money goes

A joint is a motor, a gearbox and a sensor, and a humanoid needs twenty or more of them. That sets the floor under every project here.

Berkeley Humanoid Lite is the clearest case, because the team published the line items in their paper (April 24, 2025; Robotics: Science and Systems 2025). Ten larger 6512 actuators at $188 each and twelve smaller 5010 actuators at $136 each come to $3,512 of the $4,312 total, or 81 cents in every dollar. The onboard computer is a $129 mini PC.

Poppy's repository lands in the same place: "Building a complete Poppy Humanoid costs $8000-9000 with about 60% for buying the 25 Robotis Dynamixel actuators required."

Printing the skeleton saves you the structure, which was never the expensive part. When a project quotes a low number, look at what it did to the motor count.

Berkeley Humanoid Lite: the one to start with

This is the one I would point a competent hobbyist at, because it attacks the part everyone else buys around. Rather than ordering precision metal gearboxes, the Berkeley team designed a cycloidal gearbox you print yourself. A cycloidal drive gets a large speed reduction from a lobed disc wobbling inside a ring of pins, and that rolling contact tolerates plastic far better than fine gear teeth do. Twenty-two of them, in PLA, on a machine with a 200 by 200 by 200 mm build area.

The team's own timings: the custom parts print "within a week" and the "entire robot can be assembled in about three days." Those are the designers, so read both as floors.

They published the failure mode too, which I appreciate. Across a 60-hour durability test the printed gearboxes held around 90% efficiency, while "backlash increased slightly as the 3D-printed parts experienced wear." Backlash is the free play you feel wiggling a joint with the motor off, and it is the enemy of precise control. The paper's other listed limitation: nobody has yet studied what sustained heat does to the printed structure.

Training and deployment code ships with it, and the paper reports zero-shot transfer of a walking policy from simulation to the hardware: train the gait in a simulator, load it, and it walks.

ToddlerBot: small, and already replicated by strangers

Stanford's ToddlerBot (February 2, 2025) is 0.56 m and 3.4 kg with 30 active degrees of freedom: seven per arm, six per leg, two in the neck, two at the waist. Fully 3D printed, off-the-shelf components, "keeping the total cost under 6,000 USD."

The line that matters most to a builder is the last one in that abstract. Assembly and maintenance need "basic technical expertise, as validated by a successful independent replication of the system." Somebody outside the lab built one from the files and it worked, which is worth more than any spec on the page. The docs name Bambu Lab printers specifically, for part tolerances.

Arms mean grippers, and grippers are the least-solved part of any humanoid. A knee-high 30-DOF robot will not load your dishwasher. It is the machine you teach.

The $400 tier, and what it is not

Zeroth Bot from K-Scale Labs advertises a bill of materials from $350 under an MIT license, the lowest credible entry price I found. The repo calls itself "Public Alpha" and warns to "expect breaking changes." K-Scale's own domain, kscale.dev, did not resolve for me, so treat the company as unconfirmed and the repository as the part you can rely on. Repositories outlive companies, which is one of the better arguments for open hardware.

Open Duck Mini v2 is a knee-high two-legged robot modelled on Disney's BDX droid, under $400 in parts, Apache 2.0, with an assembly guide its own repo still flags as incomplete. No arms, so a biped rather than a humanoid. For learning balance control on hardware that cannot hurt you when it falls, excellent. As a robot that does chores, it is a duck.

The old guard: InMoov and Poppy

InMoov is the ancestor, started by French sculptor Gaël Langevin in January 2012 and billed as "the first Open Source 3D printed life-size robot." Its hardware map is refreshingly concrete: two Arduino Megas, about 4 kg of natural ABS filament, a 6V 60-amp power supply, braided fishing line for the tendons, around thirty hobby servos, and any printer with a 12 by 12 by 12 cm bed.

InMoov does not walk. The legs are printable but listed as not motorized, so the build is a life-size head, torso and arms on a stand. The project publishes no total either, so every "$1,500 InMoov" figure online comes from somebody else's build rather than from Langevin.

Poppy, out of France's Inria, is the most institutionally supported of the group at 83 cm and 25 Dynamixels, and visibly a mid-2010s design: its robot page still lists a board running Ubuntu 14.04.

The license is part of the spec

If you plan to sell anything you make, read the license first. InMoov's files carry a "Creative Commons - Attribution - Non-Commercial license," and ToddlerBot's designs are CC BY-NC-SA 4.0: free to build for yourself, closed to a side business. Berkeley Humanoid Lite, Zeroth Bot and Open Duck Mini are permissive (MIT, MIT, Apache 2.0). Poppy is CC BY-SA hardware and GPL v3 software, with the name trademarked.

What you need before you start

• A printer with at least a 200 mm cube of build volume, well tuned. Tolerances decide whether printed gearboxes bind, which is why ToddlerBot's docs name a brand.
• Soldering and wiring. Every project here hides a loom of motor and CAN wiring behind its panels.
• Linux and Python, for the onboard control stack and the training frameworks.
• Simulation patience. The build is not finished when the last screw goes in. Training a walking policy and transferring it eats most of the calendar.

Build or buy

Berkeley Humanoid Lite: $4,312 in parts, ten days of your own work at the very least, 0.8 m tall. A Unitree R1 is from $4,900 assembled, shipping and customs on top, adult-sized, and it arrives working. A Unitree G1 is $13,500. Our full price list has the rest of the ladder. If you want a robot, buy the robot.

Build one when you want what a purchase cannot give you: every CAD file, every gain in the controller, the freedom to redesign a hip and print it that afternoon, and a repair bill measured in filament. Berkeley published all of it because commercial hardware "remains high-cost, closed-source, and non-transparent." The source files are the product.

If you want open files without the sourcing, Reachy Mini from Pollen Robotics and Hugging Face ships as a kit: $399 USB Lite, $499 wireless, lead time up to 90 days. It is a 28 cm desk robot with a 6-DOF head rather than a humanoid, so it sits on the shelf covered by our desk companions guide.