UC Berkeley researchers have published the full blueprints for “Berkeley Humanoid Lite,” a mid-scale robot that undercuts six-figure commercial humanoids by nearly two orders of magnitude. The team says anyone with a desktop 3D printer and a parts budget below $5,000 can build the 0.8-meter-tall biped (about 2 feet 7.5 inches tall).
Commercial humanoids such as Unitree’s H1 hover around the $90,000 mark before import fees, and Boston Dynamics and Agility Robotics have never publicly quoted anything substantially under six figures. (The Unitree H1 humanoid robot is listed at $90,000 on Unitree’s official store with a note to contact the manufacturer for the real price.) Roboworks lists the H1 at $120,135, excluding duties.
Parts for a complete Humanoid Lite add up about $4,312 in the U.S. and $3,236 in China, with ten high-torque 6512 actuators at $188 apiece and a dozen lighter 5010 units at $136 each, according to Table III of the arXv paper on the robot. Every non-standard part fits inside a 200 mm×200 mm×200 mm print envelope, so a hobby-grade FDM printer will do the job. All CAD, firmware and training scripts sit on GitHub under a permissive license.
The project has drawn some skepticism on Reddit. Critics point out that while significantly cheaper, the current capabilities of the $5,000 open-source robot don’t come anywhere close to high-end commercial systems, which offer far greater strength, speed, and dexterity. One commenter dismissed the current iteration as comparable to “toys” built a decade ago, adding that while progress in the field is inevitable, “this robot is not proof of what is possible” yet. Others question whether the $5,000 price tag truly equates to “owned by everybody” accessibility, given both the cost and the technical skills likely required for assembly and operation.
In many senses, however, the technology stack is comparatively off-the-shelf. An Intel N95 mini-PC drives four 1 Mbps CAN 2.0 buses at 250 Hz, while a 6-cell 4,000 mAh Li-Po delivers about 30 minutes of untethered runtime. The project has NSF backing (grants 2303735 and 2238346) and a disclosed stake in the Robotics & AI Institute.
In lab tests, a reinforcement-learning controller trained entirely in simulation booted on the real robot and walked without manual retuning, a coveted “zero-shot” transfer milestone for bipedal AI. Demo videos also show the bot writing its initials, manipulating blocks and noodling with a Rubik’s Cube.
The team introduces a “performance-per-dollar” ratio: average peak joint torque divided by robot size, then normalized by price. By that yardstick, the 5-grand bot beats several commercial platforms that cost an order of magnitude more
[From Berkeley Humanoid Lite]
Tell Us What You Think!
You must be logged in to post a comment.