Tightly-Coupled LiDAR-IMU-Wheel Odometry with Online Calibration of a Kinematic Model for Skid-Steering Robots

TL;DR

This study presents a tightly-coupled LiDAR-IMU-wheel odometry algorithm incorporating online calibration of a kinematic model for skid-steering robots and proposes a full linear wheel odometry factor, which not only serves as a motion constraint but also performs the online calibration of kinematic models.

Abstract

Tunnels and long corridors are challenging environments for mobile robots because a LiDAR point cloud should degenerate in these environments. To tackle point cloud degeneration, this study presents a tightly-coupled LiDAR-IMU-wheel odometry algorithm with an online calibration for skid-steering robots. We propose a full linear wheel odometry factor, which not only serves as a motion constraint but also performs the online calibration of kinematic models for skid-steering robots. Despite the dynamically changing kinematic model (e.g., wheel radii changes caused by tire pressures) and terrain conditions, our method can address the model error via online calibration. Moreover, our method enables an accurate localization in cases of degenerated environments, such as long and straight corridors, by calibration while the LiDAR-IMU fusion sufficiently operates. Furthermore, we estimate the uncertainty (i.e., covariance matrix) of the wheel odometry online for creating a reasonable constraint. The proposed method is validated through three experiments. The first indoor experiment shows that the proposed method is robust in severe degeneracy cases (long corridors) and changes in the wheel radii. The second outdoor experiment demonstrates that our method accurately estimates the sensor trajectory despite being in rough outdoor terrain owing to online uncertainty estimation of wheel odometry. The third experiment shows the proposed online calibration enables robust odometry estimation in changing terrains.

Figures (13)

• Figure 1: Indoor experiment conditions and odometry estimation results: (a) Degenerated point clouds were obtained in long-term because (b) the narrow FOV LiDAR (Livox AVIA) pointed to walls of long straight corridors for imitating severe point cloud degeneration such as tunnels. (c) All wheel radii of the skid-steering were enlarged by $25~\%$ (i.e., large model error) for verifying adaptability of our online calibration method. (d) The proposed method (Ours) was the most accurate of all cases thanks to online calibration and online uncertainty estimation although this long straight corridor includes the two areas where point clouds were degenerated and unavailable. The robot traveled about 120m, 302s. Therefore, the proposed method can tackle long-term point cloud degeneration, point cloud absence, and large kinematic model errors.
• Figure 2: System overview of the proposed method. We jointly solve LiDAR-IMU-wheel odometry and online calibration of full linear model (kinematic model for skid-steering robots) such that all those constraints are consistent. ${\bm x}_t$ is a frame state including IMU pose ${\bm T}_t = [{\bm R}_t | {\bm t}_t] \in SE(3)$, velocity ${\bm v}_t \in \mathbb{R}^3$, and bias ${\bm b}_t = [{\bm b}_t^a, {\bm b}_t^{\omega}] \in \mathbb{R}^6$. ${\bm K}_t \in \mathbb{R}^6$ is the parameters of the full linear model. The proposed method estimates ${\bm x}_t$ and ${\bm K}_t$ in real-time by using measurement sources as in LiDAR point cloud, IMU measurements (acceleration, angular velocity), and wheel encoder values of all wheels.
• Figure 3: Diagram for kinematics of skid-steering robots. Note that the result of wheel odometry described in the robot frame is finally transformed into the IMU frame. The robot velocity $[v_{x}~v_{y}~{\psi_{z}}]^\top$ is calculated by wheel angular velocities ${\bm \omega} = [\omega_{\rm L} ~\omega_{\rm R}]^\top$ and the $J$ matrix.
• Figure 4: Accuracy validation of the full linear model. We can infer that the official Rover robotics driver uses the ideal differential drive model for expressing the wheel odometry of the skid-steering robot.
• Figure 5: Used testbed for experiments in Section \ref{['subsec:indoor_environment_exp']}, \ref{['subsec:outdoor_environment_exp']}, \ref{['subsec:trans_environment_exp']}. The narrow FOV LiDAR was placed toward the lateral direction and used for odometry estimation for imitating severe point cloud degeneration (e.g., tunnels). Note that the omnidirectional FOV LiDAR was used for only obtaining reference (ground truth of the estimated trajectories) in Section \ref{['subsec:trans_environment_exp']}.