Sanim09/rover-terrain-vibration

Rover Terrain Vibration

A small, fully simulated benchmark for terrain recognition from vibration under distribution shift. A 1/10-scale car (PyBullet's racecar) drives over five procedurally generated ground types; a simulated phone-grade IMU records how it shakes. Each row is one 1-second window at 100 Hz (accelerometer xyz, gyroscope xyz, wheel-encoder speed) labelled with the ground type and the measured roughness in millimetres.

The point is the test splits: each one changes exactly one thing the model never saw.

split	what's different from train	use it to measure
test_id	new terrain layouts only	baseline accuracy
test_speed_gap	speeds 0.95-1.15 m/s (a gap inside the training range)	interpolation in speed
test_speed_fast	speeds 2.0-2.6 m/s (training tops out at 1.6)	extrapolation in speed
test_rougher	severity 0.8-1.0 (training tops out at 0.65)	extrapolation in roughness

Splits are assigned per drive, never per window, so overlapping windows from one drive never leak across splits.

Quick start

from datasets import load_dataset
import numpy as np

ds = load_dataset("Sanim09/rover-terrain-vibration")
row = ds["train"][0]
x = np.stack([row[c] for c in ["acc_x", "acc_y", "acc_z", "gyro_x", "gyro_y", "gyro_z", "wheel_speed"]], 1)  # (100, 7)
row["terrain"], row["roughness_mm"]

Columns

column	meaning
acc_x/y/z	specific force in the car frame (x forward, z up), m/s², reads +9.81 on z at rest
gyro_x/y/z	angular rate in the car frame, rad/s
wheel_speed	rear-wheel encoder speed, m/s (differs from true speed when wheels slip)
terrain, terrain_id	smooth / gravel / cobble / washboard / rocky (0-4)
roughness_mm	RMS ground height under the wheel tracks during this second, wavelengths > 1 m removed
severity	the terrain generator's roughness knob, 0-1
speed_cmd, speed_true	commanded and true mean forward speed, m/s
run_id, window	drive identifier and window index within the drive (50 % overlap)
terrain_params	JSON with the generator parameters for that drive

Raw 100 Hz recordings of every drive are in runs/ (one .npz per drive, plus runs.csv), if you want other window lengths.

Baselines

Naming the ground type - macro-F1 (chance 0.20), 95 % bootstrap interval over drives:

model	seen speeds, new layouts	speed gap 0.95-1.15 m/s	faster 2.0-2.6 m/s	rougher, severity 0.8-1
Random forest, vibration features	0.90 (0.87-0.92)	0.85 (0.81-0.87)	0.27 (0.22-0.32)	0.69 (0.65-0.73)
Random forest, + wheel speed	0.91 (0.88-0.93)	0.89 (0.86-0.92)	0.29 (0.22-0.34)	0.72 (0.67-0.75)
Random forest, per-metre features	0.92 (0.90-0.94)	0.93 (0.91-0.95)	0.80 (0.76-0.84)	0.75 (0.72-0.78)
Linear, vibration features	0.88 (0.85-0.89)	0.89 (0.87-0.91)	0.57 (0.51-0.62)	0.78 (0.74-0.80)
Linear, + wheel speed	0.91 (0.88-0.93)	0.93 (0.91-0.95)	0.73 (0.68-0.77)	0.76 (0.72-0.79)
Linear, per-metre features	0.88 (0.86-0.90)	0.91 (0.88-0.93)	0.70 (0.65-0.75)	0.72 (0.67-0.75)
MLP on spectra + speed	0.89 (0.87-0.91)	0.87 (0.85-0.88)	0.57 (0.52-0.62)	0.77 (0.73-0.80)
MLP on raw signal + speed	0.78 (0.75-0.81)	0.71 (0.67-0.74)	0.47 (0.42-0.51)	0.64 (0.60-0.66)
1-D CNN on raw signal + speed	0.93 (0.91-0.95)	0.95 (0.94-0.96)	0.57 (0.51-0.63)	0.73 (0.69-0.75)

Feature sets: vibration statistics; + wheel speed; per-metre (spectra converted to bumps per metre using the wheel speed). Full roughness results, code and the speed-normalisation features: https://github.com/sanim-science/rover-terrain-vibration.

Headline: every model does well at seen speeds (0.88-0.93), but at faster speeds a random forest falls to 0.27 and a 1-D CNN to 0.57, while per-metre features keep 0.80. If you have a method that is robust to speed shift, test_speed_fast is the split to beat.

How it was made, and what we checked

Simulator: PyBullet 3.25, physics at 1000 Hz, rigid contact, wheel collision shapes replaced by spheres of the same radius. We picked this setup because it was the only one that reads ~0 vibration on perfectly flat ground; the stock cylinder wheels, soft contact, or physics at 100-400 Hz all invent 1-9 m/s² of fake vibration there, which a classifier would happily learn. Features also agree within ~6 % between 1000 Hz and 2000 Hz physics. Washboard ridges of spacing L at speed v produce a peak at v / L Hz (within 4 %), as they should. Full code and checks: https://github.com/sanim-science/rover-terrain-vibration.

Limitations

Simulation only. The car has no suspension and rigid wheels; real tyres, suspension, frame resonances, IMU mounting, temperature drift and sensor clipping are not modelled. Treat this as a controlled benchmark for how models behave under shift, not as a stand-in for real rover data.

Citation

Report: coming soon.