Kart Telemetry Experiment

Can a camera mounted on a kart's steering wheel recover the driver's steering input?

This experiment uses an Insta360 Go 3S gyroscope and an AiM MyChron logger to separate wheel rotation from chassis yaw, align the recordings, and produce per-lap plots. It grew out of a practical question: could existing onboard recordings reveal something useful about driving technique?

Status: concluded as an exploratory steering study. The processing pipeline ran on real track data. Absolute steering accuracy and a benefit to driving performance were not established. The code and results remain here as a case study; no further capture or feature work is planned for this experiment.

Real-session results · Executed notebook · Run the example · Validation limits

A lap from the experiment

XTreme, 11 May 2026. The steering trace is a filtered estimate; the other panels come from the logger. This demonstrates the pipeline's output, not independent steering accuracy.

The historical run produced six GPS-segmented lap windows. The camera/logger alignment used a time offset of −14.03 s and reached a normalized correlation of 0.658. That correlation describes the alignment signals, not steering accuracy. See the results for the inputs and all six plots.

The engineering problem

A wheel-mounted gyro rotates with both the steering wheel and the kart. Integrating it directly mixes those two motions. In the simplified model:

gyro projected onto column = steering rate + k × chassis yaw rate
k = |column axis · vertical|

The real-session path estimates the column axis from gyro covariance, uses quiet accelerometer samples to estimate its tilt, and aligns the camera with the logger by correlating yaw-related signals. It subtracts the projected logger yaw rate, integrates the remainder, and removes slow integration drift.

flowchart LR
    A[Wheel-camera IMU] --> C[Axis and tilt estimate]
    A --> D[Time alignment]
    B[MyChron GPS yaw] --> D
    C --> E[Subtract projected chassis yaw]
    D --> E
    E --> F[Integrate and filter]
    F --> G[Per-lap steering estimate]

Loading

This path requires the logger. Geometry estimation assumes wheel rotation dominates the relevant gyro motion, and the yaw model does not account for every chassis movement. Filtering suppresses drift but also suppresses sustained steering, so the output is not an absolute steering-angle sensor.

Run the example

No camera, logger, private recordings or telemetry decoder is needed for this deterministic synthetic example. From a clone of this repository, with Python 3.11 or newer:

python3 -m venv .venv
.venv/bin/python -m pip install -r examples/requirements.txt
.venv/bin/python -m examples.steering_demo --out-dir results/demo

Open results/demo/steering-demo.png. The command also writes the input and output traces to CSV and a JSON summary. The checked-in version is below:

The example sends generated motion through the existing steering_from_synced function. Axis, tilt, bias and clock alignment are supplied as known inputs. It demonstrates the subtraction/integration/filter stage and its limitation on a held steering angle. It does not validate camera decoding, automatic geometry, synchronization or accuracy on a kart.

Explore real recordings

The executed notebook includes the historical plots and can be read on GitHub without running it. It retains the historical implementation snapshot and outputs; _build_notebook.py generates a new, unexecuted notebook from the current modules. Raw recordings are not distributed, so the historical result cannot be regenerated from a clone alone.

To process your own compatible recordings:

.venv/bin/python -m pip install -r pipeline/requirements.txt
.venv/bin/python -m analysis.per_lap \
  /path/to/wheel.gyroflow /path/to/session.xrk \
  --all --out-dir results/my_session

The .gyroflow file must contain the embedded IMU samples. Original Insta360 video input additionally requires the external gyro2bb decoder. In the Go 3S recordings examined here, Pro Video preserved raw gyro and ordinary Video did not. Preserve originals; re-encoded video is not a substitute for sensor data.

What the experiment established

• A working analysis path: camera IMU ingestion, clock alignment, geometric projection, logger-yaw subtraction and lap plots were exercised on real data.
• A repeatable mathematical demonstration: the synthetic example makes the signal separation visible while showing how drift filtering changes the result.
• A clear evidence boundary: there is no completed comparison against a calibrated, independent steering reference, and no demonstrated coaching or lap-time improvement. Visual plausibility is not calibration.

The project stops here because the next step would be a separate measurement validation effort, while the original goal was better karting. Preserving the method and its limits is a useful outcome without that further commitment.

Explore the repository

Artifact	Purpose
RESULTS.md	Historical real-session output and interpretation
analysis.ipynb	Executed walkthrough with embedded plots
examples/steering_demo.py	Runnable synthetic illustration
pipeline/sync_xrk.py	Clock alignment and geometry estimation
analysis/per_lap.py	Steering estimate and lap visualization
STEERING_VALIDATION.md	What remains unverified
ROADMAP.md	Closeout and disposition of the original scope

The tests exercise decoding, calibration, synchronization and analysis contracts with synthetic fixtures. After installing the full requirements, run:

.venv/bin/python -m pytest -q

Experimental tire/vision utilities retained in the repository are outside this steering case study and are not presented as validated capabilities.