Uc Tekerlek Kinematikli Mobil Robotlarin Cevrim Ici Odometri Kalibrasyonu
Authors/Creators
Description
The odometry of mobile robots with tricycle kinematics or single-steered vehicles is calculated using parameters derived from manuals or CAD drawings. However, parameters such as the wheel radius may change due to friction or payload, or manufacturing tolerances may cause the wheelbase and steering encoder offset to deviate from nominal datasheet values. These variations systematically introduce erroneous measurements into localization filters or controllers relying on odometry. In this study, five parameters used in wheel odometry estimation — the wheel radius (r), the wheelbase (w_b), the lateral offset of the drive wheel (d), and the forward/backward steering offsets (α_f and α_b) — are estimated online using an MSCKF (Multi-State Constraint Kalman Filter)-based multi-sensor state estimation filter. Experiments show that the lateral offset of the drive wheel (d) absorbs the lateral offsets of other sensors. Since no two-dimensional planar motion can decouple these parameters, the lateral offset is examined as an effective lateral offset, defined as d_eff = d + y_O - y_L, where y_O and y_L denote the lateral mounting positions of the wheel and LiDAR frames. Testing with synthetic sensor simulations under 20 different erroneous initializations demonstrates that the wheel radius estimation error is reduced to 0.04 ± 0.07 mm (compared to an initial error of 2.7 mm). A consistency audit conducted across ten independent noisy sequences clears four parameters while identifying that only the wheelbase contains a deterministic model error. Second, physics-engine-based simulation tests show that the proposed calibration trajectory achieves superior results compared to the UMBmark benchmark. Furthermore, the benefits of calibration are demonstrated over a 30-second LiDAR outage by comparing the performance of the calibrated filter against the uncalibrated filter. A mid-operation change of the wheel radius under payload is shown to be re-converged within 10–30 seconds through a random-walk parameter model, and the measured computational cost shows that the five-parameter calibration adds negligible overhead (the lidar-free stack runs in 0.5% of the data duration) while LiDAR matching dominates.
Files
161.pdf
Additional details
Additional titles
- Translated title (English)
- Online Odometry Calibration of Mobile Robots with Tricycle Kinematics