- The paper introduces a novel LiDAR-inertial framework with a CUDA-accelerated frontend and EKF backend to achieve high-speed, low-latency odometry.
- It employs sparsity-aware GICP and adaptive submaps to reduce sensor noise effects and mitigate motion-induced distortions under extreme dynamics.
- Experimental results show 38.4% reduced runtime and 6.9% improved accuracy, outperforming state-of-the-art methods on diverse datasets.
FAR-LIO: Fast, Accurate, and Robust LiDAR-Inertial Odometry for High-Speed Autonomous Systems
Overview and Motivation
The paper "FAR-LIO: Enabling High-Speed Autonomy through Fast, Accurate, and Robust LiDAR-Inertial Odometry" (2606.26010) addresses the persistent challenge of real-time, high-accuracy odometry under extreme dynamics and sensor noise, particularly within high-speed autonomous racing. Existing LiDAR-Inertial Odometry (LIO) methodologies often suffer from substantial latency, sensitivity to parameter tuning, and limited robustness in highly dynamic scenarios, limiting their applicability in latency-constrained environments such as autonomous vehicles operating at speeds up to 250 km/h. FAR-LIO is developed to directly confront these limitations, introducing a CUDA-optimized architecture designed for low-latency operation without compromising accuracy across diverse environments and sensor configurations.
Algorithmic Architecture
CUDA-Accelerated Frontend
The core innovation in FAR-LIO is its LiDAR scan pipeline, which exploits a CUDA-based voxel hashmap (cuVoxelMap) for spatial data structuring, facilitating parallelized nearest-neighbor search and efficient map updates. This is complemented by a sparsity-aware Generalized Iterative Closest Point (GICP) algorithm with adaptive thresholding, leveraging robust statistical kernels and covariance regularization (via Frobenius norms) for enhanced robustness under sparse or noisy data distributions. The pipeline incorporates:
- Motion Undistortion: Parallel correction of LiDAR scan distortion based on IMU velocity histories.
- Adaptive Local Submaps: Dynamic adjustment of voxel densities and sizes according to environmental point densities, maintaining scalable memory and runtime characteristics.
- Sparsity-Aware Registration: Handling regions with limited geometric information through fallback point-to-point registration, ensuring alignment robustness.
Extended Kalman Filter Backend
Odometry output from the LiDAR pipeline is fused at 100 Hz with high-frequency IMU data using a Kinematic Extended Kalman Filter (EKF). Notably, the EKF incorporates delay compensation and measurement upsampling, addressing the temporal asynchrony typical between LiDAR and IMU sensors. Calibration of IMU biases is performed offline to enhance robustness in high-noise operational regimes. The EKF also stabilizes roll and pitch estimates using reference angles, mitigating z-drift and long-term accumulation errors.
Experimental Evaluation
Datasets and Setup
FAR-LIO is exhaustively evaluated across diverse datasets including KITTI, MulRan, and proprietary autonomous racing sequences from the Indy Autonomous Challenge and Abu Dhabi Autonomous Racing League. The evaluations encompass multiple sensor modalities (Velodyne, Ouster, Luminar LiDARs; varied IMUs) and environments (urban, highway, racetrack), with runtime and accuracy benchmarks conducted on the autonomous racecar’s compute platform (Intel Xeon D-2166NT, NVIDIA RTX A5000) under ROS.
Computational Efficiency
FAR-LIO achieves significant reductions in runtime, outperforming the next best baseline (Faster-LIO) by an average of 38.4%. It maintains computation times well below supported LiDAR frequencies (9.59 ms on open-source datasets, 19.23 ms on racing data with 200,000-point scans), demonstrating scalability and real-time capability. Unlike CPU-centric methods, the CUDA-based design yields tightly clustered and predictable callback latencies, critical for robust closed-loop control and preventing odometry divergence caused by occasional latency spikes.
Odometry Accuracy
FAR-LIO consistently delivers superior accuracy—with an average positional error (APE) and relative positional error (RPE) reduced by 6.9% compared to state-of-the-art baselines using a uniform parameter set. The method outperforms others on high-speed racing data, KITTI, and MulRan, completing all sequences robustly where several baselines failed (e.g., KISS-ICP diverged under extreme dynamic regimes, Faster-LIO and Fast-LIO2 incomplete on KITTI's low-frequency IMU sequences). The adaptive submap density and motion undistortion modules reduce both local inconsistency and long-term drift.
Ablation and Latency Studies
Ablation experiments confirm the critical contribution of each architectural module: EKF with delay compensation, motion undistortion, sparsity-aware GICP, and adaptive submap density together induce significant reductions in both APE and RPE. Latency simulations demonstrate that odometry error remains stable up to a dataset-dependent threshold (60–100 ms), beyond which divergence ensues. FAR-LIO’s runtime reliably operates well within this domain, underscoring its robustness for real-world deployment.
Implications and Future Directions
FAR-LIO’s open-source, GPU-accelerated architecture sets a new standard for LIO frameworks aimed at high-speed, real-time applications. Its ability to generalize across environments and sensor types with a single parameter set substantially lowers the operational complexity and risk for deployment in challenging autonomous contexts (e.g., racing, urban navigation). The demonstrated robustness against latency-induced instability highlights the importance of architectural design for real-time odometry pipelines, rather than reliance solely on improved scan registration accuracy.
Theoretically, FAR-LIO motivates further investigation into hybrid frontends/backends for odometry, adaptive spatial data structures, and temporal compensation methods. Practically, its extensibility suggests utility not only in high-speed vehicles but in aerial robotics, SLAM for large-scale industrial environments, and multi-agent systems requiring low-latency pose updates. Emergent trends such as FPGA/GPU co-design, distributed SLAM, and fault-tolerant sensor fusion could integrate FAR-LIO’s modular design for even broader robustness and performance gains.
Conclusion
FAR-LIO demonstrates a significant advance in LiDAR-Inertial Odometry, combining a CUDA-accelerated, sparsity-aware GICP frontend with an EKF backend optimized for delay-compensated sensor fusion. The framework achieves state-of-the-art accuracy and efficiency on both public and proprietary datasets, robustly supporting high-speed autonomous operation. Its generalizability across sensor configurations and environments, validated through rigorous empirical analysis and simulation, establishes it as an authoritative reference for future research and development in real-time LIO systems.