Skip to content

Browser Calibration

Open the browser calibration page

The page calibrates an IMU against a trajectory of another sensor. It runs entirely in your browser: Calibrex, NumPy, and SciPy run under Pyodide, and no data leaves the page. It estimates:

  • the rotation of T_lidar_imu;
  • the clock offset (t_imu = t_sensor + dt);
  • the gyro bias;
  • the lever arm, from the accelerometer.

The computation, held-out evidence, and artifacts are the same as for the MID360 IMU-LiDAR benchmark. The only difference is that the sensor motion comes from a trajectory you already have, instead of Calibrex's own LiDAR odometry.

Looking for calibrex check or calibrex estimate instead? The bag check page plans, checks and estimates on your own rosbag2 in the browser; see the workflow and Plan and run a bag in the browser.

Inputs

Input Format
IMU CSV t, gx, gy, gz, ax, ay, az per line: seconds, rad/s, and m/s² or g. Commas or spaces; a header line is allowed.
Sensor trajectory TUM: t x y z qx qy qz qw per line, poses T_world_sensor, # for comments. For example, a LiDAR odometry or SLAM output.

Both inputs must share one clock, up to the constant offset that is calibrated. The trajectory is cut into 10-second windows, and a gap longer than 0.5 s starts a new block. At least three windows are needed, and a few minutes of motion that rotates about every axis gives the best results.

Press Use a synthetic example to see a hand-held recording with a known truth being calibrated. It takes about 10 s to load Python and about 10 s more to calibrate.

Outputs

  • Verdicts. The rotation and the lever arm each get a pass, warn, fail, or inconclusive verdict, with the reasons.
  • Per-quantity table. Each quantity has its reported std, its status (estimated or unobservable), and whether held-out windows detected the known-bad control.
  • Extrinsic. The 4×4 T_lidar_imu and the clock offset.
  • Artifacts. Downloadable, schema-valid slac.imu_lidar_rotation/v0.1 and slac.imu_lidar_translation/v0.1 artifacts with provenance, including the SHA-256 of your inputs.

Checking the page

tools/check_browser_page.mjs runs the exact Python snippets embedded in the page under Pyodide in Node. See the script header for how to run it.

The same thing on the command line

calibrex imu-lidar trajectory --imu imu.csv --trajectory odometry.tum \
  --output-rotation rotation.yaml --output-translation translation.yaml
calibrex validate translation.yaml

Use --acceleration-unit g for Livox drivers, and --mid360-reference to compare the result with the Livox MID360 design extrinsic.

Limitations

  • The trajectory is not checked. If it comes from a LiDAR-inertial odometry, it already depends on the IMU and on the extrinsic that odometry assumed.
  • Motion must excite every axis. Planar or yaw-only motion leaves axes unobservable, and the verdict says so.
  • Accelerometer scale and axis misalignment are not estimated.