Calibration¶

Calibration is optional. Pelorus produces a usable trajectory on the defaults with a well mounted sensor pair. Calibrating the IMU removes drift that tuning cannot.

Dedrunk does the fitting. It reads MCAP recordings and ROS 2 bags and writes Kalibr format YAML. It ships with pelorus[all] and the Homebrew package, and installs on its own with cargo install dedrunk. Its documentation covers the motion each fit needs, how to read the results and what to do when a fit refuses. This page covers the Pelorus side: recording and applying the results.

The steps can be done in any order. The extrinsic fit is cleaner with the intrinsics already applied.

Recording¶

pelorus online raw runs the native sensor drivers without the odometry and records their output unchanged. It publishes on the normal topics while recording, so the rest of your stack keeps working.

# Both sensors, for the intrinsics and the extrinsics
pelorus online raw --output-rosbag calib_drive

# IMU only, for the noise measurement
pelorus online raw imu --output-rosbag imu_static

One drive with both sensors covers intrinsics and extrinsics if it follows Dedrunk's recording guide. The noise measurement needs its own recording of at least three hours. dedrunk topics <bag> lists what was recorded. --output-mcap works as well.

For the intrinsics, dedrunk imu record listens while pelorus online raw imu publishes, prompts for each orientation, and fits the recording once all are captured:

pelorus online raw imu &
dedrunk imu record /ouster/imu --bag calib_intrinsics.mcap -O imu_intrinsics.yaml

IMU intrinsics and noise¶

dedrunk intrinsics imu calib_drive -O imu_intrinsics.yaml
dedrunk imu noise imu_static -O imu_noise.yaml
dedrunk imu merge imu_intrinsics.yaml imu_noise.yaml -O imu.yaml --pelorus imu.rl

--pelorus imu.rl writes the Pelorus side of the result as Ratslang: imu.calibration.parameter_file pointing at imu.yaml, and the four imu.covariance_* values. Include it from the main configuration instead of copying values over:

= ./imu.rl

imu.calibration {
    parameter_file = "./imu.yaml"

    # Measure the gyroscope bias at start-up instead of trusting the file.
    gyroscope = true
    gyroscope.samples = 100
}

Pelorus applies the scale, misalignment and bias in the file to every IMU sample. Keep the robot still for the first second after start-up while the gyroscope bias is measured.

dedrunk imu noise also prints the four imu.covariance_* values derived from the measurement, ready to paste, and writes them with --pelorus PATH. The defaults are tuning values, not a noise model of any sensor, so the measured ones are optional. They come from a still sensor, and Dedrunk inflates them by noise.inflation (10 by default, -o noise.inflation=20 to change it) to account for vibration and temperature changes on a moving robot. Compare trajectories on a recording you know before keeping them.

Lidar to IMU extrinsics¶

The extrinsic fit needs a trajectory produced without the IMU, since a lidar-inertial trajectory already assumes an extrinsic:

pelorus offline calib_drive -c pelorus.rl -o imu.lidar_only=true -o output.mcap.rosbag_dir=lidar_traj
dedrunk extrinsics lidar-imu lidar_traj --imu-bag calib_drive --pelorus lidar_imu.rl

imu.lidar_only ignores the IMU and imu.from_lidar_transform. Roll and pitch are relative to the start pose in this mode, so use it only for calibration.

Read Dedrunk's observability report before accepting the output. A level loop drive does not constrain every parameter, and an existing calibration is better than an unconstrained replacement. Dedrunk prints a block with imu.from_lidar_transform and imu.calibration.time_offset to paste into the configuration, and --pelorus lidar_imu.rl writes the same block to a file to include with = ./lidar_imu.rl. Includes resolve relative to the including file, and a setting assigned after an include overrides it.

imu.calibration.time_offset is how far the IMU clock runs ahead of the Lidar clock. Pelorus subtracts it from every IMU timestamp. The effect grows with turn rate. At 1 rad/s, 10 ms of skew rotates the scan by about half a degree.

Refining online¶

With a good starting extrinsic, from CAD or from Dedrunk, the filter can refine it while running:

imu.from_lidar_transform.estimate_online = true

It needs a close initial guess and enough excitation. On a robot that mostly drives straight it drifts, so compare the result against the starting value.