BERKELEY HUMANOID LITE: AN OPEN-SOURCE HUMANOID ROBOT WHOSE ACTUATORS ARE 3D PRINTED

A walking humanoid robot whose joints are printed, not bought, and the whole machine lands under five thousand dollars.

by Hybrid Robotics group, UC Berkeley

FULL CAD BOM FIRMWARE DOCS

RoboticsOpen-hardware

Built withSTM323D printing

difficulty
●●●●●
time
a few weeks
license
MIT
repo
repo ACTIVE1,953 stars
1
Jump to section

COMPAREE VERDICT

Berkeley Humanoid Lite is a research platform published by the Hybrid Robotics group at UC Berkeley. It is 80 cm tall, weighs about 16 kg and moves on 22 modular actuators, each a hobby brushless drone motor driving a 3D-printed cycloidal gearbox, with an off-the-shelf motor driver and a magnetic encoder. The paper prices the larger 6512 actuator at about 188 dollars and the smaller 5010 at about 136 in US parts, and the whole robot at about 4,300 dollars. The project publishes CAD, print files, the BOM, motor-controller firmware, reinforcement-learning training code and robot descriptions in three formats; the code is MIT and the other assets CC BY-SA 4.0, so commercial use is allowed. This is not a weekend project: the paper says the custom parts print within a week and the robot assembles in about three days, but it assumes you can assemble precision gearboxes, flash and configure 22 motor controllers, and train and debug a walking policy. The actuators are where things go wrong — the paper itself names gear backlash as the dominant source of position error — so build and test one before committing to all 22.

GOOD TO KNOW

  • —CAD for the robot and both actuators is on Onshape, and the print files are Bambu Lab MakerWorld projects, all linked from the documentation's Releases page.
  • —Bill of materials is a Google sheet in the official docs, with Amazon and Taobao purchase links.
  • —Firmware for the actuators and robot control stack is in the repository, plus a reinforcement-learning training pipeline.
  • —Robot descriptions ship in URDF, MJCF and USD (Universal Scene Description) formats for simulation and control.
  • —Code is MIT licensed; CAD, documentation and other assets are Creative Commons Attribution-ShareAlike 4.0, permitting commercial use with attribution and share-alike.
  • —Step-by-step build docs (BOM, tools, print settings, actuator and robot assembly) are on the official documentation site — written for a robotics lab rather than a first-time builder.

Parts to buy

9 items

From our check of the build. Exact quantities and part numbers are in the creator’s BOM.

  • 22 brushless drone motorsFind
  • 22 B-G431B-ESC1 motor driversFind
  • 22 AS5600 magnetic encodersFind
  • Bearings and fastenersFind
  • Intel N95 mini PCFind
  • Four USB-CAN adaptersFind
  • BNO085 IMU and a 6S LiPoFind
  • Paper's US total is about 4Find
  • 300 dollarsFind

BUILDS OF THE WEEK

Five open-source builds worth your weekend, every week.

Checked like this one: what’s really in the repo, what it costs, how hard it is. One email, unsubscribe anytime.

Can I build this?

PrintActuator housings, cycloidal disks and shafts for 22 actuators, plus the robot's structural parts, in standard PLA. Any desktop printer with at least a 200 x 200 x 200 mm build volume works; the docs give tuned print profiles for the Bambu Lab X1C.
Buy22 brushless drone motors (10 for the larger 6512 actuators, 12 for the 5010), 22 B-G431B-ESC1 motor drivers, 22 AS5600 magnetic encoders, bearings and fasteners, an Intel N95 mini PC, four USB-CAN adapters, a BNO085 IMU and a 6S LiPo. The paper's US total is about 4,300 dollars.
ToolsDesktop 3D printer, soldering iron (a heat-insert stand is recommended), hot glue gun for mounting the encoder magnets, hex keys from M2 to M6, screwdrivers, STM32CubeIDE for flashing the motor controllers, and a Linux workstation with an NVIDIA GPU for training.
SkillsAdvanced. You must understand brushless motor control, embedded firmware, reinforcement learning pipelines, and mechanical assembly of precision gearboxes. Prior experience with legged robots or actuator builds is effectively required.
TimeAbout a week to receive parts, about a week to print the custom parts and about three days to assemble the robot, according to the paper. Flashing 22 motor controllers, calibrating the actuators and getting a policy to walk on your robot come on top of that.
Cost$$$. The paper's published hardware total is about 4,300 dollars at US prices (about 3,200 at Chinese prices); that excludes the printer, tools, failed prints and a GPU workstation for training.
SafetyLithium battery pack under load during gait testing. Actuators under power can pinch or crush; secure the robot during initial testing. No mains voltage in the robot itself.

Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.

Videos

Berkeley Humanoid Lite: An Open source, Accessible, and Customizable 3D printed Humanoid Robot.

The official project video from the Hybrid Robotics lab; the build docs also link video tutorials for assembling the actuator.

More builds like this

All projects

Gallery

opengraph.githubassets.com
img.youtube.com

Start here

Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.

  1. 1.Read the paper (arXiv 2504.17249 covers the design rationale, actuator tests and BOM cost tables; the purchase links are in the BOM sheet in the documentation.)
  2. 2.Download the CAD and print files (The documentation's Releases page links the CAD on Onshape and ready-made print projects on Bambu Lab MakerWorld for the robot and both actuators.)
  3. 3.Order motors and encoders(The BOM in the paper specifies exact motor models and encoder part numbers. Order all 22 sets before you start printing — lead times vary.)
  4. 4.Print and assemble one actuator first(Build and test a single actuator completely before committing to all 22. This is where print tolerances and bearing fits will show up.)

Resources

Documentation, files and community threads for this build — we link straight to the original sources and never rehost the creator’s files.

KNOWN ISSUES

  • Cycloidal gearboxes are sensitive to print accuracy; the paper names gear backlash as the main source of position error. Print one actuator, test it under load and measure backlash before printing the remaining 21.
  • The official docs have step-by-step assembly guides for the actuator and the robot, but they assume lab-level skills. If you have never built a gearbox, follow the actuator video exactly and expect a learning curve.
  • Training a stable gait requires a Linux workstation with a GPU and familiarity with reinforcement learning pipelines. The provided training stack is not a one-click solution — you will tune hyperparameters and wait hours for convergence.
  • The motor controllers run the project's own open-source firmware (Recoil BESC), which you flash and configure per joint with STM32CubeIDE — budget time for getting the CAN IDs and motor parameters right.
  • The mechanical design is documented, but the control stack assumes you already know how to stabilise a biped and interpret IMU data in real time.
  • Lithium battery packs under continuous actuator load get hot. Budget for a battery management system and thermal monitoring, neither of which are in the base BOM.

Can I use different motors or encoders?

The actuator firmware and gearbox geometry are designed around the specified motor dimensions and encoder mounting. Substituting parts means redesigning the housing CAD and retuning the control loop — possible, but not a weekend task.

How long does it take to print all 22 actuator housings?

The docs do not give a total print time. Each actuator is several printed parts (housing, cycloidal disks, shafts, spacers) printed with two tuned profiles, and there are 22 actuators plus structural parts, so plan for days of printing and some reprints.

Do I need a simulation environment to test this?

Yes. The project ships robot descriptions in URDF, MJCF and USD (Universal Scene Description) formats, plus Isaac Lab training and a MuJoCo sim-to-sim check. Test policies in simulation before running them on the real robot.

Is there a smaller or simpler version to start with?

There is no smaller kit, but the actuators are modular: the paper shows the same joints rebuilt as a biped, a quadruped, a centaur-like robot and a mobile base, and the repo includes biped (legs-only) policies. To de-risk, build one actuator and test it before committing to the full humanoid.

Can I sell robots built from this design?

The code is MIT licensed and the CAD is CC BY-SA 4.0, so yes, commercial use is permitted. You must provide attribution and share modifications to the CAD under the same licence. Verify compliance with both licences before selling.

Community builds

No community builds yet — be the first, we feature the best ones.

Discussion1

FROM THE COMPAREE TEAM

The whole machine lands under five thousand dollars, and the paper prices the larger 6512 actuator at about 188 dollars in US parts and the smaller 5010 at about 136. If you were sourcing this, would you print all 22 actuators yourself or buy commercial units for the critical joints and print the rest?

CompareeTEAM2mo agoedited

Practical notes from our verification: the repository holds the training, sim-to-sim and deployment code, and the hardware side lives in the project documentation on GitBook — a materials and parts list with Amazon and Taobao links, step-by-step sections for building the actuator, flashing the motor controllers and building the robot, plus an official video tutorial for actuator assembly. The actuators are the entire point of the project: drone motors driving 3D-printed cycloidal gearboxes in two sizes, which is how the paper keeps the whole robot under 5,000 dollars in US prices. Print accuracy still matters — a cycloidal gearbox is sensitive to dimensional errors, and the paper itself measures backlash as the main source of position error. Build one actuator first, test it, and check backlash before committing to all 22. Code is MIT; the other assets, including the CAD, are CC BY-SA 4.0, so commercial use is allowed with attribution and share-alike. Correction (4 October 2026): we re-checked this page line by line against the project's own repository, documentation and videos, and fixed errors in earlier versions.

Hybrid Robotics group, UC Berkeley

The Hybrid Robotics group at UC Berkeley, led by Koushil Sreenath, builds legged robots and control algorithms for dynamic locomotion. Berkeley Humanoid Lite was published by Yufeng Chi, Qiayuan Liao, Junfeng Long, Xiaoyu Huang, Sophia Shao, Borivoje Nikolic, Zhongyu Li and Koushil Sreenath as an open-source, low-cost platform to make humanoid research accessible to groups priced out of commercial robots.

GitHub

Star the project on GitHub

DISCLAIMER

  • Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
  • Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
  • Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.