Berkeley Humanoid Lite

Berkeley Humanoid Lite is an open-source humanoid robot research platform developed at the University of California, Berkeley. The 2025 peer-reviewed system is 0.8 m tall, weighs 16 kg, and has 22 actuated joints plus two grippers. Its structure and actuator gearboxes are made largely with desktop 3D printing, while motors, bearings, controllers, sensors, and fasteners are commercially sourced. The team presented the platform at Robotics: Science and Systems XXI in June 2025.^[1][2]

The project was designed to make a mid-scale robot easier to fabricate, modify, and repair without a machine shop. That goal should be separated from the measured evidence. The paper reports a component bill of materials below $5,000, a set of actuator tests, a learned walking controller, and operator-controlled manipulation demonstrations. It does not establish a universal delivered price, a complete service-life rating, or autonomous manipulation. Berkeley Humanoid Lite also was not the first open-source humanoid to use printed parts. Poppy and NimbRo-OP2 were documented in peer-reviewed work in 2014 and 2017, respectively.^[1][11][12]

Publication and design scope

Yufeng Chi, Qiayuan Liao, Junfeng Long, Xiaoyu Huang, Sophia Shao, Borivoje Nikolic, Zhongyu Li, and Koushil Sreenath authored the paper "Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot." The official proceedings assign it DOI 10.15607/RSS.2025.XXI.062. Project documentation records an initial public release on April 28, 2025.^[1][5]

The published robot is one arrangement of a modular actuator system. Its paper illustrates other possible arrangements, including quadruped, centaur-like, mobile-base, and lengthened-leg concepts. Those figures demonstrate how the same joint modules could be rearranged; they are not performance evaluations of every illustrated morphology. An adult-height drawing with carbon-fiber link extensions likewise is a design example, not a tested second model.^[1]

The platform belongs to a broader line of open hardware intended for research and education. Poppy showed rapid morphological exploration with a 3D-printed humanoid, and NimbRo-OP2 described an adult-size open platform with a printed exterior and gears. Berkeley Humanoid Lite's distinguishing engineering choice is a pair of modular cycloidal drive actuators built around widely available brushless DC motors, rather than a claim to have originated open-source or printed humanoids.^[1][11][12]

Hardware architecture

The paper and released training guide give the following baseline configuration.^[1][9]

Scroll sideways for more →

The 22-joint count comes from the released training environment and matches the paper's ten 6512 plus twelve 5010 actuator modules. The two grippers appear as separate rows in the bill of materials. Counts of 24 sometimes arise when grippers are treated as additional actuated devices, but that is not the convention used by the full-body training task.^[1][9]

The larger 6512 actuator uses an MAD Components M6C12 150KV motor. The 5010 uses a smaller 5010 motor. Both designs use a B-G431B-ESC1 motor driver and an AS5600 magnetic encoder, together with bearings, fasteners, printed housings, shafts, and cycloidal disks. The joints are self-contained, so a designer can alter link lengths or joint ordering without redesigning a remote transmission.^[1]

The paper gives three reasons for selecting printed cycloidal reductions. Their load is shared across multiple contacts, the geometry can tolerate some of the resolution limits of fused-deposition printing, and a reducer can be produced with printed parts plus ordinary bearings and steel hardware. Earlier experiments by Roozing and Roozing showed that low-reduction printed cycloidal gearboxes can be viable, while also documenting how geometry, material, run-in, play, friction, and stiffness affect results. Those studies support the design class, but their test articles and conditions differ from Berkeley's actuators.^[1][13][14]

Berkeley's structural parts use PLA. Through-screws and embedded brass hex stands reinforce printed components across weak layer directions, and some subassemblies are combined into single prints to reduce fasteners and stress concentrations. Every custom part in the paper fits a 200 x 200 x 200 mm build volume. The torso's aluminum extrusions provide mounting space for the computer, battery, IMU, and interface electronics.^[1]

The four limb buses divide traffic and wiring. Each operates at 1 Mbps using CAN 2.0, and USB-CAN adapters connect them to the N95 computer. The IMU reaches the computer through an Arduino and USB. The paper reports an approximately 30-minute operating time from the 6S battery, but it does not provide a standardized workload, payload, or battery-aging protocol for that figure. External power is supported for longer stationary tests.^[1]

Author cost estimate and fabrication

The paper provides separate United States and China bills of materials, with each value rounded to the nearest dollar. The complete totals are $4,312 and $3,236. These are the authors' 2025 component estimates, not a sale price or a promise that another builder will reproduce the total.^[1]

Scroll sideways for more →

Within those totals, one 6512 actuator was estimated at $188 in the United States and $157 in China; one 5010 was $136 and $94. The authors said most general parts were available from Amazon or Taobao/AliExpress and that electronic parts came from distributors including DigiKey and Mouser.^[1]

The estimate omits important parts of a delivered build cost. The paper does not add the purchase or depreciation of a printer, soldering and heat-insert tools, shipping, tax, import duties, failed prints, builder labor, calibration time, safety equipment, spares, or maintenance. Its performance-per-dollar figure also combines the team's own component estimate with published selling prices for some comparison robots. It is therefore an author-defined screening metric, not a controlled comparison of equivalent capabilities or lifetime cost.^[1]

The authors estimated that most stocked parts could arrive within one week, printing could finish within another week, and final assembly could take about three days. Current instructions make the skill boundary clearer. A builder needs a desktop printer, soldering iron, heat-insert equipment, hot-glue gun, and hand tools. Print profiles were tuned for a Bambu Lab X1C, and the documentation warns that other printers can require changes. The actuator build also requires moving resistors and soldering wires on an encoder board, mounting a radial magnet on the motor shaft, wiring the controller, and calibrating each unit.^[5][6][7][8]

These requirements make it a fabrication project rather than a kit assembled from finished modules. Availability and price also vary by country and date. The paper supports a sub-$5,000 component estimate under the authors' sourcing assumptions, not a universal build price.^[1]

Actuator evaluation

The authors tested the actuator modules on a custom dynamometer at 24 V, using the same proportional-derivative gains and bandwidth settings as on the robot. A second actuator held speed while two load cells measured output torque; a separate board measured supply voltage and current. They also cross-checked results from their controller and firmware against a Moteus controller.^[1]

The paper distinguishes gearbox mechanical efficiency from total actuator efficiency. Mechanical efficiency was measured output power divided by commanded torque times velocity. Total efficiency was measured mechanical output divided by electrical input, so it also included motor copper loss and driver loss. The gearbox reached about 90% mechanical efficiency across much of the tested range, with efficiency falling at high torque and speed. This result should not be restated as 90% total electrical efficiency.^[1]

A static test fixed the output of a 6512 unit and ramped commanded torque in both directions. A linear fit over 4 to 10 Nm gave an estimated transmission stiffness of 319.49 Nm/rad. The paper compared that value with a different printed cycloidal reducer made from carbon-fiber-reinforced polyamide, but differences in material and geometry prevent a direct platform ranking.^[1][14]

For durability, a 6512 actuator repeatedly lifted a 0.5 kg pendulum with a 0.5 m arm from -45 to +90 degrees at 0.5 Hz for 60 hours. Testing paused hourly for the first 12 hours and every 12 hours afterward to measure efficiency and backlash. Efficiency first decreased and then returned near its original value; backlash increased as the printed parts wore. The authors judged the backlash acceptable in that experiment. This is one specified laboratory duty cycle, not a general service-life or shock rating for the joint or robot.^[1]

The team also evaluated fabrication variation. Six 6512 samples printed on two machines were tested at 1 rad/s, and their torque error stayed within plus or minus 0.5 Nm over the reported range. Six newly printed 6512 modules had a maximum observed backlash of 0.0229 rad and a standard deviation of 0.0042 rad. A five-joint arm then reached four targets 100 times each under SteamVR tracking; the end-effector position standard deviation was 3.433 mm. That last number is repeatability in the cited setup, not absolute positioning accuracy throughout the workspace.^[1]

The paper identifies thermal behavior as an unresolved limitation. It did not study prolonged heating of the printed structure deeply enough to establish how temperature affects strength and reliability. Later maintainer notes add that the printed actuators were too fragile for high-performance tasks and that controller connectors and cables could fail during extended use. Those qualifications are important context for the shorter benchtop tests.^[1][4]

Control and demonstrations

Learned locomotion

The walking controller used reinforcement learning. The team trained a PPO policy in Isaac Gym as a partially observed control problem. Its observations included base angular velocity, projected gravity, joint positions and velocities, a user-commanded linear velocity, and the previous action. The policy returned desired joint positions.^[1]

The onboard multilayer perceptron ran at 25 Hz, distinct from the 250 Hz actuator and IMU communication rate. The paper reports direct sim-to-real transfer of the walking policy without an additional state estimator and shows forward walking while following a user velocity command. It also says the experiment limited the actuators to 30% of their torque setting. That percentage applies to the demonstrated controller; it is not a general payload reserve, safety factor, or proof that a scaled-up morphology will work unchanged.^[1]

The current software guide uses NVIDIA Isaac Lab with Isaac Sim 4.5.0 and Isaac Lab 2.1.0 on a tested Ubuntu 24.04 setup. It defines one velocity task for all 22 joints and another for the 12 leg joints. The guide estimates about two hours for 6,000 training iterations but does not identify the training GPU on that page, so the duration is not a portable performance benchmark. The repository also provides robot descriptions for MuJoCo, sim-to-sim tools, configuration files, and ONNX policy checkpoints.^[3][9]

Teleoperated manipulation

The manipulation demonstrations used teleoperation, not an autonomous task planner. Two SteamVR base stations and hand controllers tracked the operator. Pink and Pinocchio supplied inverse kinematics for the two five-joint arms, and the operator controlled the grippers.^[1]

The paper describes a third-person "headless" mode and a first-person virtual reality mode. In the first, controller motion is interpreted in a global frame while the operator watches the robot. In the second, controller changes are mapped between the operator's and robot's local frames. The reported demonstrations include writing with a marker, packing and unpacking a box, picking and placing blocks, and manipulating a Rubik's Cube.^[1]

Those images establish dexterous, human-directed motion with the cited tracking and inverse-kinematics stack. They do not show that the robot independently perceived a cube state, planned a solution, or executed the sequence without an operator. Likewise, the paper's education and animatronics sections are proposed uses, except for its specific report that a 6512 actuator was already used in UC Berkeley's MECENG 102B course.^[1]

Open-source release and versioning

The versioned repository contains training environments, simulation and sim-to-sim entry points, real-robot deployment code, low-level interfaces, motion-capture and teleoperation tools, robot-description submodules, configuration files, and pretrained ONNX checkpoints. The v1.1.0 README states that repository code is licensed under the MIT License and "other assets" under Creative Commons Attribution-ShareAlike 4.0.^[3]

That split should not be summarized as one license covering every item in a build. External dependencies, submodules, vendor firmware, purchased components, and trademarks retain their own terms. The MIT license also disclaims warranty. Public CAD, software, and documentation make inspection and modification possible, but do not make the robot a certified product.^[3]

Project documentation records the initial release on April 28, 2025. The v1.1.0 GitHub release was published on Sept. 7, 2025 and consolidated fixes and additional instructions. The documentation log records corrected 6512 print files, swapped bearing descriptions, a missing bearing and insert, added wiring and calibration guidance, replacement of an incorrect 5010 housing, addition of a missing limit stop, and later correction of joint limits. Reproduction therefore depends on using current assets and instructions rather than only the original paper or an early download.^[4][5]

Reproducibility and safety

The release covers several layers needed for reproduction, but completeness of files is different from independent replication. The authors publish hardware designs, bills of materials, firmware and control code, training tasks, checkpoints, and step-by-step build documentation. Their release notes also show community members building and modifying the platform.^[3][4][5]

An external data point comes from the EPFL AI Team's SAPIEN project. The team says its active humanoid project begins from Berkeley Humanoid Lite. This is evidence that another engineering group adopted the design as a starting point. It is not a peer-reviewed reproduction of the original $4,312 cost, 60-hour actuator test, 3.433 mm repeatability result, or learned walking experiment.^[15]

No standardized multi-site replication study of those quantitative results was identified by Aug. 21, 2026. Public build reports also expose practical variation in print tolerances, soldering, wiring, magnet placement, per-motor calibration, and parts revisions. The project's own instructions warn that the CAN pads on the motor controller are fragile and that the system includes high-power electronics with no guarantees.^[8][10]

The paper's reported test boundary is also narrower than a complete humanoid qualification program. It does not provide drop testing, fall recovery, payload envelopes, water or dust protection, battery-cycle aging, electromagnetic-compliance testing, thermal lifetime, or a full field-failure distribution. Builders and researchers must establish their own operating limits and safety procedures.^[1][4][10]

Later development

The Berkeley Humanoids organization now describes the original repository as the initial Berkeley Humanoid Lite paper codebase. Its other repositories should not be used to silently revise the specifications above.^[16]

Berkeley Humanoid Lite Arm, labeled BHL V1.5, is a separate bimanual upper-body platform. Its README replaces the original custom M6C12 and 5010 actuator modules with Robstride actuators, adds LeRobot-compatible robot and teleoperator plugins, and says the project remains under active development with likely breaking changes. It also notes that actuator selection and the head design were still being refined.^[17]

The organization separately indexes V2-oriented robot-description and control repositories. That public repository index shows an ongoing development direction, not a completed independent validation of a full V2 robot. The 2025 full-body platform and its V1 evidence therefore remain the subject of this article.^[4][16][17]

Funding and competing interest

The paper acknowledges support from National Science Foundation award 2303735 for POSE, NSF award 2238346 for CAREER, and the Robotics and AI Institute. It also states that Koushil Sreenath has a financial interest in the Robotics and AI Institute and that he and the organization may benefit from commercialization of the research. The other project materials do not convert the paper's cost and accessibility claims into independent evaluations, so the disclosure should be considered alongside those claims.^[1][2]

See also

• Robotics

References

1. ^Chi, Yufeng, et al. "Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot." Robotics: Science and Systems XXI. 2025. DOI 10.15607/RSS.2025.XXI.062. roboticsproceedings.org/...p062
2. ^UC Berkeley Hybrid Robotics. "Berkeley Humanoid Lite." Accessed Aug. 21, 2026. lite.berkeley-humanoid.org
3. ^Hybrid Robotics. "Berkeley Humanoid Lite," repository at v1.1.0. Sept. 7, 2025. github.com/...v1.1.0
4. ^Hybrid Robotics. "Berkeley Humanoid Lite v1.1.0 Release." Sept. 7, 2025. github.com/...v1.1.0
5. ^Berkeley Humanoid Lite Docs. "Releases." Release log through Nov. 10, 2025. berkeley-humanoid-lite.gitbook.io/...releases
6. ^Berkeley Humanoid Lite Docs. "Preparing the Tools." Accessed Aug. 21, 2026. berkeley-humanoid-lite.gitbook.io/...ing-the-tools
7. ^Berkeley Humanoid Lite Docs. "3D Printing Instructions." Accessed Aug. 21, 2026. berkeley-humanoid-lite.gitbook.io/...-instructions
8. ^Berkeley Humanoid Lite Docs. "Building the Actuator." Accessed Aug. 21, 2026. berkeley-humanoid-lite.gitbook.io/...-the-actuator
9. ^Berkeley Humanoid Lite Docs. "Training Environment." Accessed Aug. 21, 2026. berkeley-humanoid-lite.gitbook.io/...g-environment
10. ^Berkeley Humanoid Lite Docs. "Home." Accessed Aug. 21, 2026. berkeley-humanoid-lite.gitbook.io/...home
11. ^Lapeyre, Matthieu, et al. "Rapid morphological exploration with the Poppy humanoid platform." IEEE-RAS International Conference on Humanoid Robots. 2014. DOI 10.1109/HUMANOIDS.2014.7041479. hal.science/...document
12. ^Ficht, Grzegorz, et al. "NimbRo-OP2: Grown-up 3D Printed Open Humanoid Platform for Research." IEEE-RAS International Conference on Humanoid Robots. 2017. DOI 10.1109/HUMANOIDS.2017.8246944. ais.uni-bonn.de/...Humanoids_2017_NOP2.pdf
13. ^Roozing, Wesley, and Glenn Roozing. "3D-printable low-reduction cycloidal gearing for robotics." IEEE/RSJ International Conference on Intelligent Robots and Systems. 2022. DOI 10.1109/IROS47612.2022.9982006. ris.utwente.nl/...cloidal_gearing_for_robotics.pdf
14. ^Roozing, Wesley, and Glenn Roozing. "Experimental comparison of pinwheel and non-pinwheel designs of 3D-printed cycloidal gearing for robotics." IEEE International Conference on Robotics and Automation. 2024. DOI 10.1109/ICRA57147.2024.10610250. ris.utwente.nl/...cloidal_gearing_for_robotics.pdf
15. ^EPFL AI Team. "SAPIEN." Accessed Aug. 21, 2026. epflaiteam.ch/...sapien
16. ^Berkeley Humanoids. Organization profile and repository index. Accessed Aug. 21, 2026. github.com/Berkeley-Humanoids
17. ^Berkeley Humanoids. "Berkeley Humanoid Lite Arm (BHL v1.5)." Accessed Aug. 21, 2026. github.com/...Berkeley-Humanoid-Lite-Arm