Mars
rover
wheel

no. 2

Rovers read terrain by eye, and sand fools them. This scaled Curiosity wheel feels the ground instead: strain gauges on tapered spokes classify terrain from its vibration. Validated with NASA JPL engineers.

Team of three.

86%
terrain accuracy
10
terrains, Hyde Park
ESP32
strain gauges + ADC
Design
Three spoke profiles in FEA. Tapered wins on frequency separation.
Test
Single-spoke sweep with an audio exciter, then field runs.
Result
Random Forest classifies terrain and flags slippage.
The problemThe ideaSpoke designOne spoke on the benchWhat the spoke saidThe full wheelFrom signal to patternResult
1 / 9

The problem

Rovers judge the ground by camera. A camera can’t feel softness, and sand that looks fine can trap a wheel.

2 / 9

The idea

So let the wheel feel it. Every surface flexes the spokes differently, and strain gauges can read that.

3 / 9

Spoke design

Three spoke shapes, simulated. The tapered one kept the vibration frequencies furthest apart.

4 / 9

One spoke on the bench

One printed spoke, clamped and shaken through a frequency sweep to check the simulation.

5 / 9

What the spoke said

Each gauge picks up its own frequency range, so two gauges beat one.

6 / 9

The full wheel

A 60% scale printed wheel on a portable rig. Four strain channels, about 1000 readings a second.

7 / 9

Out in the park

Ten surfaces: concrete, cobbles, grass, gravel, soft sand and more.

8 / 9

From signal to pattern

Each wheel turn is split into frequency bands. The energy in each band is that surface’s fingerprint.

9 / 9

Result

A Random Forest learns the fingerprints. It mostly mixes up look-alike surfaces, and wheel slip shows clearly.

Watch

Build and field-test montage

Try it

Each surface shakes the wheel's spokes in its own way. Split that vibration into frequency bands and the pattern is a fingerprint a classifier can learn. Pick a terrain.

strain on one spoke over one wheel turnenergy per frequency band

Classifier reads this as . Illustration only: the shapes are simplified, not recorded data.