Tofu Hardness Analyzer

A desktop-scale hardness measurement device built from a salvaged 3D printer.

Full machine at a three-quarter angle with a tofu sample loaded on the build plate — Press play to watch a full test cycle. The video streams here without leaving the page.

Overview

A desktop-scale measurement device that determines the hardness of test samples, built to showcase mechatronics skills and to learn image recognition. The machine identifies tofu pieces on its build plate using computer vision, then performs a compression test on each one to classify it as soft, firm, or extra firm.

At a glance

Base: Salvaged 3D printer
Controller: ESP32-S3 Nova Cam
Motion: 3× MKS42 closed-loop
Sensing: HX711 load cell
Vision: OpenCV on PC
Metric: Effective modulus (kPa)
Output: Soft · Firm · Extra firm

How It Started

I bought a used 3D printer off Marketplace intending to harvest it for parts. After pulling the stepper motors and the motherboard, I decided it would be more interesting to turn it into something else instead.

At the time I was applying for a mechanical engineering position at a food startup, so I built a hardness analyzer for tofu. The goal was to take real measurements and determine whether a sample was soft, firm, or extra firm.

The donor 3D printer as purchased, before conversion — Before: donor printer.

The finished hardness analyzer after conversion — After: the analyzer.

Hardware Decisions

Closed-loop stepper motors

MKS42 servo

Each module integrates the driver and encoder directly onto the motor. This simplified wire management considerably and gave better positional precision than the open-loop drivers the printer shipped with.

Controller + camera

ESP32-S3 Nova Cam

Chosen for two reasons: enough RAM to handle image capture and buffering, and a built-in camera, which is one less subsystem to integrate.

Homing

3× limit switches

Kept the printer's three original limit switches to home all axes and establish a known origin on startup.

Force sensing

Custom HX711 load-cell mount

Replaced the print head with a custom-designed holder for an HX711 load cell, which is how the machine reads contact and compression force.

Custom mount with the HX711 and load cell installed on the carriage — Load-cell assembly.

One MKS42 servo mounted on the frame, showing the integrated driver and encoder — MKS SERVO42C Closed-Loop Stepper.

The ESP32-S3 Nova Cam controller board with its built-in camera module — ESP32-S3 Nova Cam mount.

System Architecture

The ESP32-S3 Nova Cam sits at the center of the build, driving motion, reading force and limit switches, and passing frames to the PC over USB serial for vision processing.

Detailed wiring diagram: ESP32-S3 WROOM connected to three stepper motors and drivers (X, Y, Z), three limit switches, an HX711 amplifier and load cell, a 3.3 V rail, and a 12 V 20 A power supply. — Full wiring diagram, the actual pin-level connections for the build.

How It Works

1

Homing

On startup, the machine homes all three axes against the limit switches to establish a known origin.
2

Imaging

The build plate moves to the camera position and the ESP32 captures an image.
3

Vision processing

An OpenCV pipeline on the PC detects four fiducial markers that define the test area, counts the tofu pieces inside it, and computes each piece’s center coordinates.
4

Positioning

The ESP32 drives the X and Y axes to move the first sample under the probe.
5

Contact & height

The Z axis descends while the load cell is read continuously. The height at which force first rises marks contact, and gives the piece’s true height for free.
6

Compression curve

The probe presses a further 3 mm at a slow, controlled rate, streaming the force every 0.1 mm to capture the full force-vs-displacement curve.
7

Stiffness → class

The curve becomes an effective compressive modulus (kPa), a size-independent stiffness that grades the piece soft, firm, or extra firm.
8

Repeat

The machine indexes to the next piece and repeats until every sample in the test area has been measured.

Process flow: contact detection loops until the load threshold is met, then each piece indexes back to positioning.

Processed camera frame with four fiducial markers highlighted, detected tofu pieces outlined, and center points marked — OpenCV output: markers define the test area; each detected piece is outlined with its center marked.

Measuring Stiffness, Not Force

The machine grades each piece by its stiffness, an effective compressive modulus in kilopascals. Stiffness is a property of the tofu itself, so it stays the same whether a piece is cut thick or thin.

That independence is the point: tofu comes in all thicknesses, and a thinner slice compresses more under the same press than a thick one. Reading stiffness rather than raw push force lets the classifier compare pieces of any size on equal footing. The machine captures the full force-vs-compression curve on every press to compute it.

Stress

kPa

How hard the probe pushes, spread over its 426 mm² tip.

force ÷ probe area

Strain

ratio

How far the piece squished, relative to its own height.

compression ÷ height

Modulus

kPa

Stiffness, the same for a big or small cube of one tofu.

stress ÷ strain

Soft below 16.1 kPa · firm up to 30 kPa · extra firm above. Thresholds set from three tofus of known firmness, and refined as more are tested.

Results

The machine reliably locates multiple samples in a single pass and classifies each one against calibrated stiffness (modulus) thresholds without operator input beyond loading the plate.

Future Improvements

Next, I plan to generate stress-strain curves from the load-cell data using MATLAB. After that I'll likely convert the machine into a plotter, since I don't have an ongoing need for a hardness analyzer. I built this to learn image recognition and to prove I could take a project like this from salvaged parts to a working system. It served both purposes.