A tiny open-source biped robot that walks on six of the cheapest servos you can buy. No ankles, no servo driver: an ESP32-C3, a gyroscope and six SG90s. He walks by rocking from foot to foot, and stepping with whichever foot is free.
Everything to build him is here: the parts to print, the wiring, the firmware, and the page you calibrate and drive him from. There's also a simulation you can drive in your browser.
| Folder | What's inside |
|---|---|
print/ |
Every part to print, in print orientation, with quantities and settings |
docs/wiring.png |
The wiring diagram |
firmware/ |
The ESP32-C3 sketch, and how to flash it |
control/ |
Calibrate and drive him from your browser (Python, runs on your computer) |
docs/calibration.md |
Calibration, step by step |
cad/v1.3/ |
STEP files and the FreeCAD assembly, to edit the design |
dist/, mujoco/ |
The browser simulation and a native MuJoCo viewer |
| Part | Qty |
|---|---|
| SG90 9 g servo | 6 |
| ESP32-C3 SuperMini, with a shield/expansion board that has servo headers | 1 |
| GY-521 gyroscope (MPU-6050) | 1 |
| 1.3" I2C OLED (SH1106) | 1 |
| 5 V buck converter | 1 |
| Flat LiPo battery, to fit the cassette under the body (about 65 × 55 × 10 mm) | 1 |
| M2 screws, wires | — |
| PLA | about 200 g |
Everything is in print/, with what to print first and the settings.
| Joint / module | GPIO |
|---|---|
| Left hip roll, hip pitch, knee | 3, 4, 6 |
| Right hip roll, hip pitch, knee | 0, 1, 2 |
| Gyroscope SDA / SCL | 8 / 9 |
| OLED SDA / SCL | 10 / 7 |
All grounds are common. The servos take 5 V from the regulator, never from the board's 3.3 V pin.
See firmware/: one library install, then compile and upload.
The step nobody talks about. Every servo horn sits up to half a tooth off, and that is enough to stop
him from walking. A guided page fixes it in five minutes: docs/calibration.md.
pip install pyserial
python3 control/duckbert_web.pyPress Wake up, then hold ▲. More in control/.
He can't stand on one foot: without ankles, tilting the hips moves his weight less than 3 mm over the
foot. So he doesn't try. The gyroscope feels him rocking side to side, and the hips push the rock
along like a swing. The rule is in step_control() in the firmware: when he's leaning onto his front
foot, and still moving that way, the back foot is free, so it steps. Then the other foot is in front,
and the same rule fires on the other side.
The curved shoes help: he rolls onto a foot instead of tipping over its edge.
A standalone simulation of an earlier version (v1.2, flat feet), with its own walking gait and directional controls in the browser. MuJoCo simulates gravity, contacts and the six servomotors; Three.js renders the CAD meshes. It has no connection to the physical robot.
- W / ↑: walk forward.
- S / ↓: walk backward by reversing the phase of the same gait.
- A / ← and D / →: rotate left or right in place.
- Space: pause the simulation.
- With a mouse or touchscreen, tap a direction to keep moving; use Pause to freeze the simulation.
- Reset to centre: restore the starting position, orientation and gait phase.
- Drag to orbit the camera; scroll or pinch to zoom.
Directions are relative to the robot. Left and right rotate in place; combine them with W or ↑ for a walking turn. Pausing freezes simulation time; it is not a physical stopping controller. The simulation also pauses when the page loses focus.
The simulator imports the neutral assembly meshes from cad/out_none_v1.2,
including the shorter body, LiPo cassette, compact hip block, ESP32-C3 shield and inward-offset
feet. The top colour button switches between Bordeaux red (#800020) and a white body
with orange feet and cyan frame, brackets and horns. Both options keep the six servos blue,
the screen black and the eyes white. The ground has a fixed white grid at 10 cm spacing.
Masses, centres of mass and inertias are derived from those meshes with solid PLA density (1240 kg/m³), the CAD manifest's module mass estimates, and 22 g of wiring and fasteners. These remain estimates, especially battery weight and printed infill. The flat 54 × 41 mm sole contacts are centred at ±32.5 mm. Joint limits come from the v1.2 manifest. Ground collisions use convex hulls for the printed parts and boxes for the soles; link-to-link collisions are not simulated.
The single v11_walk gait uses the existing none_oled periodic controller, selected
after testing it on the imported model. Its steering sign is calibrated to this gait.
The controller implementation is ported from gait9.py, with position commands at 50 Hz, torque limits
and command slew limits. Steering varies the relative hip swing amplitude by up to 10%.
Backward movement reverses the same periodic trajectory. The controller does not directly
set the free body's position or orientation, or apply artificial steering forces.
The automated check runs 12 seconds of forward, left, right and backward movement in the WASM engine, checking that the robot stays upright and turns in the expected direction. These are nominal checks, not a stability guarantee for every sequence: prolonged turns, changes of direction or numerical differences can cause falls. Backward movement may drift.
This flat-sole, no-ankle geometry is also the wall the walk runs into: see
docs/walking-limits.md for what that costs it, in millimetres, and
the two unbuilt changes with the most headroom.
Requires Node.js 22+ and Python 3 for the static server.
npm ci
npm run vendor
npm test
npm run serveOpen http://127.0.0.1:8768. The dist/ folder includes all runtime dependencies and can also
be served by any static HTTP server. No API keys or backend are required.
The repository also includes a native MuJoCo runner and English setup instructions in
mujoco/README.md. Install the Python dependency, then run
python mujoco/run_viewer.py to open the v1.2 model in MuJoCo's interactive viewer.
The main branch contains source code, tests, data and pinned dependencies. The gh-pages
branch contains only the contents of dist/, served at the project site's root.
git add .
git commit -m "Update simulation"
git push origin main
git subtree split --prefix dist -b pages-release
git push origin pages-release:gh-pages
git branch -D pages-releaseIn Settings → Pages, select the gh-pages branch and the / folder.
Imports and assets use relative paths compatible with the project URL.
MuJoCo's threaded WebAssembly runtime requires cross-origin isolation. On GitHub Pages, a project-scoped service worker adds the required COOP/COEP headers and reloads the page once on the first visit. It does not cache assets or handle requests outside this origin. Startup checks isolation before importing the engine and reports setup or loading failures instead of leaving the loading screen indefinitely. Use HTTPS or localhost and a browser that supports service workers and shared WebAssembly memory.
- Physics: Google DeepMind MuJoCo, official
@mujoco/mujocopackage 3.6.0, Apache-2.0; licence indist/vendor/mujoco/LICENSE. - Rendering: Three.js 0.178.0, MIT;
licence in
dist/vendor/three/LICENSE. - SG90 geometry and controller: the author's local Microduck project. The CAD reconstruction references the kinematic proportions of pollen-robotics/microduck; upstream meshes are not included in this repository.
Model provenance is recorded in dist/assets/provenance.json.
With Python, NumPy and trimesh installed, run
python scripts/import-cad.py /path/to/out_none_v1.2 and then npm test.
The importer uses neutral assembly exports, copies each STL unchanged, rebuilds mass properties
and collision placement, and records mesh checksums in the provenance file.