Ph.D. student in the Department of Electrical Engineering and Computer Sciences, Yufeng Chi is part of a team of Berkeley engineers that has developed Berkeley Humanoid Lite, a low-cost, open-source robot made of 3D-printed parts. They recently presented a paper on their work at the 2025 Robotics Science and Systems conference.
Humanoid robotics has evolved rapidly over the years, with devices now being designed for applications in industrial automation, healthcare, research, and personal assistance. Despite growing interest in this field, Chi said most commercially available hardware remains relatively expensive. However, with the passage of time, cost has become the easier problem to solve. The bigger challenge, he said, has been addressing customization and resource issues.
Commercial companies may offer completely built robots, but the proprietary nature of their hardware and software design often makes it difficult to replace parts or to modify components. This limits opportunities for novice roboticists to experiment with customization and to further explore the boundaries of humanoid technologies.
At the same time, not everyone has access to state-of-the-art equipment. “Some research projects assume you have access to fancy computer numerical control (CNC) machines for advanced fabrication and are using customized printed circuit boards for the electronics,” said Chi. “In a fully outfitted lab, creating a robot from scratch may be doable, but for many, including hobbyists and DIYers, it’s just out of reach.”
By developing Berkeley Humanoid Lite, the researchers set out to provide a launching point for anyone with an interest in humanoid technology. “Our aim is to help people in the research and education space become familiar with how a humanoid system works, including how to assemble and develop a robot platform, by providing them with a template to get started,” he said. “It’s all about looking into an example system and building this system from scratch, one block after another. Then after you’ve gained experience and confidence, you can build upon it and take it to a new level.”
The core of the design is a modular 3D-printed gearbox for the actuators and robot body. All other components can be obtained from widely available sources or fabricated with standard desktop 3D printers. This keeps the total hardware cost under $5,000, a fraction of the cost of purchasing a commercially built robot of similar scale. In addition, replacements for broken or worn parts can be easily fabricated using the 3D printer.
Once completed, Berkeley Humanoid Lite stands at about one meter and weighs approximately 16 kg. Chi estimates that it would take a novice roboticist about one week to build their own Berkeley Humanoid Lite, depending on one’s skillset and experience level. “The good news is that in our Discord and other community group chats, we are seeing users actually building it,” he said. “They have displayed pictures of assembled robots, which is pretty exciting.”
Recognizing that 3D-printed parts inherently lack the strength of materials such as aluminum, the researchers adopted a cycloidal gear design for the gearbox inside the actuator. “The main benefit is that the gear’s teeth are very large,” said Chi. “This distributes the load across a larger surface area than traditional gear systems, reducing stress and wear.” In addition, they tested several aspects of the 3D-printed actuators to validate their durability. “Our findings showed that the 3D-printed actuator is at least on par with other actuators,” Chi said, “or within the tolerance margin to achieve these higher-level tasks and skills.” He added, “We designed this so that if an actuator breaks down, you can just print another gear box and swap it out, [but] we have yet to break a single actuator on any of our test robots, even after putting them through all these experiments.”
The researchers also tested Berkeley Humanoid Lite’s ability to complete simple tasks, such as grasping objects and walking forward. To enable hand manipulation, the researchers assembled a teleoperation system for the robot. Using joystick controls, they demonstrated Berkeley Humanoid Lite’s ability to grab and play with items, including a Rubik’s Cube. The researchers also used reinforcement learning to develop a locomotion controller that enables the bipedal robot to walk.
Berkeley Humanoid Lite’s hardware design, embedded code, and training and deployment frameworks are fully open source. The researchers want users to be able to see how everything works and to easily customize the robot.
