TongDou

I finally finished the “confused accountant” scene that I mentioned in the first TongDou project log.

I gave TongDou a tiny ledger and let him handle my coffee debts.

That was a mistake.

He clears one debt, invents suspicious new fees, accepts a bribe, forgets the original number, and eventually decides that I now owe him exactly 3.1415 coffees.

He also refuses to round down.

The scene is made from five small variations:

- normal payment
- invented fees
- accepting a bribe
- selective memory
- absurd inflation

The interaction is scripted and runs locally. Prerecorded voice, OLED expressions, ledger animations, sound effects, and small wheel gestures are synchronized to the same scene timeline.

I recorded my own replies in Chinese and added English subtitles. TongDou answers in English, so it feels less like a hardware demonstration and more like an argument with a tiny accountant living on my desk.

The accounting is terrible, but at least the timing works.

Before TongDou had sensors, audio, or a personality, he started as two laser-cut brass side plates and eight 1 mm brass rods.

This short video shows how the brass frame is aligned, soldered, and gradually turned into the skeleton that holds the PCB, motors, battery, radar, and OLED display.

Seven rods are soldered first. The final rod is left out until the OLED is installed, because it passes between the OLED ribbon cable and its small PCB.

It is a slightly awkward way to build a desktop robot, but the exposed brass structure is a large part of what gives TongDou his character.

TongDou has moved from the V8 prototype board to V9.

TongDou V8 prototype PCB

TongDou V9 prototype PCB 

This is not just a layout cleanup. V9 adds two new ways for the little machine to understand its own body and interact with people:

- a capacitive touch input built into the gold TongDou logo
- a QMI8658A 6-axis accelerometer and gyroscope

The logo is now more than decoration. A tap or hold can be used as a physical input without adding another mechanical button.

The IMU is mounted directly on the PCB. TongDou uses two small geared motors without wheel encoders, so wheel alignment, friction, motor differences, and the surface under him can all affect how he moves. The IMU gives me a way to measure the motion of the body itself.

The first goals are simple:

- verify the actual sensor axes
- measure turning and yaw drift
- collect repeatable motion data
- prepare for small real-time movement corrections

The PCB is installed upside down inside the brass frame, so the IMU coordinate system is remapped in firmware.

I also separated the hardware test firmware from the complete TongDou application firmware and published the test package on GitHub.

The public firmware can test the display, RGB LED, microphone, speaker, power system, touch input, IMU, motors, battery, I2C devices, and LD2412 presence radar. It also starts its own Wi-Fi access point and provides a small browser-based diagnostic page.

TongDou V9 browser-based hardware diagnostic page

The radar diagnostics were tested on a real V9 board:

- empty scene: 25/25 frames reported no target
- hand wave: 25/25 frames reported a target
- 23 frames correctly identified a moving target
- nearest distance was about 60 cm, matching the test setup
- bad serial frames: 0

If the radar module is replaced or its mounting direction changes, the empty-scene background calibration should be run again.

The public test package is here:

https://github.com/solitary-dev-50/tongdou/tree/main/hardware_test_firmware

This is hardware bring-up and diagnostic firmware, not the complete TongDou application firmware. The PCB source files, mechanical files, character behaviors, audio assets, and complete production firmware are still being prepared for a later phased release.

For now, V9 is assembled, the new sensors are responding, the radar diagnostics are working, and the hardware test tools are public.

This is the first public TongDou prototype.

The current build uses an ESP32-S3, brass frame, custom PCB, OLED eyes, audio output, LEDs, two geared wheels, sensors, and an AI-connected personality layer.

The first demo scene synchronizes prerecorded voice, OLED expressions, lights, and wheel gestures to the same audio timeline.

Next step: improving the “confused accountant” scene and documenting more of the hardware and firmware.