Open-Source Quadruped With Twelve Servos And IMU Leveling

Oct 13, 2026 Last reply: 2 hours ago 4 Replies
Jump to replies (4)

When it is running, this quadruped trots and walks around indoors, holds its chassis level on an inclined cushion using IMU data, and runs an idle breathing animation. A puppet mode matches the chassis orientation to the tilt of a handheld smartphone. The head carries a rectangular display with customizable blue digital eyes, which is a nice touch of character.



PrintedRobotics built it around an ESP32 S3 with a PCA9685 servo driver, an MPU6050 IMU, and a voltage sensor board. Two 15A adjustable buck converters feed the legs, one per pair, while a separate 5V 3A converter keeps the ESP32 isolated. Twelve DS3218 high-torque servos drive the twelve degrees of freedom, down from eighteen on the earlier hexapod this design came from. The frame, leg linkages, and head casing are printed in red, white, and black, held together by roughly a hundred M3x8 screws, four M4x30 screws, and twelve ball bearings. A 3S LiPo between 3000mAh and 6000mAh powers it, and tests show over 20 minutes of runtime.



Two details stand out to me. Sacrificial printed clips snap during impacts to protect the servo motors, which seems a smart way to make crashes cheap. Servos and the controller run on independent converters to avoid brownouts without custom PCBs, and a Wago connector plus lever-nut splices keep wiring simple.



The open question: both antennas are cosmetic only right now, since an antenna plug burned a board and a regular ESP32 S3 was fitted instead, and the benefit of running two antennas is debatable anyway. The current inverse kinematics walking looks mechanical; reinforcement learning training is planned in the hope of more fluid motion. ESP32 scripts and an Android controller apk are on GitHub, with CAD, print files, and a wiring diagram shared online. At around 200 USD, that seems a solid deal.



Project by PrintedRobotics. Watch it on their YouTube channel:

formatting link


How are the servo signals actually routed through the PCA9685 board on this one? With twelve servos and a separate buck converter per leg pair, I am curious how the wiring is laid out in the chassis and whether the lever-nut splices sit near the legs or closer to the center.

The sacrificial printed clips are the choice I would highlight. Letting a cheap clip snap instead of a servo during an impact trades a reprint for an expensive part, and on a robot meant for sim2real practice that will take knocks, that seems like a very sensible design decision.

Since the maker hopes reinforcement learning will make the walking more fluid than the current inverse kinematics motion, a concrete step could be starting that training on the trot gait first, since trot already works indoors. Building a policy on top of a gait that runs reliably over 20 minutes might show progress sooner.

On the antennas: the maker has said both are currently cosmetic only. An antenna plug burned a board, so a regular ESP32 S3 board was fitted instead. The maker also mentioned a splitter could be used, but considered the gains from two antennas debatable, so that point is still open.

Join the Discussion

Have something to add? Share your thoughts — no account required.

Didn't find your answer?

Ask the community — no account required