Technical Summary: FAR-LIO

Problem Statement

Robust and accurate odometry estimation is critical for modern robotics, yet it faces significant challenges in environments characterized by highly dynamic motion, sensor noise, and extreme speeds. Autonomous racing exemplifies these conditions, operating in unstructured settings at speeds up to 250 km/h. In such scenarios, minimizing odometry latency is essential for stable closed-loop control. While many existing LiDAR-Inertial Odometry (LIO) frameworks offer high accuracy, they often require extensive parameter tuning and substantial computational resources. Many CPU-based methods fail to meet real-time constraints on target hardware, particularly when dealing with high-density point clouds or varying environmental conditions. Furthermore, few approaches explicitly address the impact of computation latency on system robustness in real-world, high-speed applications.

Methodology

FAR-LIO is a highly optimized, CUDA-accelerated LiDAR-inertial odometry framework designed to achieve Fast, Accurate, and Robust performance. The system architecture consists of two primary components: a LiDAR Scan Pipeline and a Sensor Fusion module.

1. LiDAR Scan Pipeline

This pipeline processes incoming LiDAR scans using extensive GPU acceleration:

• Preprocessing & Undistortion: Scan distortion is corrected using a linear regression model of linear and angular velocities derived from the sensor fusion state. This continuous motion model is evaluated in CUDA, allowing parallel processing of all points to project raw points to the scan's final timestamp.
• GICP Registration: The core registration utilizes a novel CUDA-based voxel hashmap (cuVoxelMap) with adaptive density. This data structure, inspired by the iVox paradigm, employs a static map with linear probing for hash collision resolution. It supports parallelized nearest-neighbor (kNN) searches.
  • Sparsity-Aware GICP (SA-GICP): The system employs a Generalized Iterative Closest Point algorithm that adapts to point sparsity. If a point lacks sufficient neighbors for valid covariance estimation, the algorithm defaults to a point-to-point registration mode.
  • Adaptive Thresholding: To reject outliers, the system uses an adaptive thresholding scheme based on the deviation between the initial guess and the registered pose.
  • Robust Kernel: A Cauchy kernel is used for least-squares optimization, with its scale dynamically adjusted.
• Adaptive Local Submap: The map is maintained as a local submap around the vehicle using the cuVoxelMap. It employs Adaptive Submap Density (ASMD), where the voxel capacity is adjusted based on the local point density to reflect the environment accurately. Voxels beyond a 1000 m radius are removed to constrain map size and ensure fast update times.

2. Sensor Fusion

The backend utilizes an Extended Kalman Filter (EKF) running at 100 Hz to fuse asynchronous IMU data with LiDAR odometry:

• Kinematic Model: The filter uses a purely kinematic motion model with state variables for position, orientation, and velocity. IMU biases are calibrated during standstill rather than estimated online to ensure robustness in high-noise environments.
• Delay Compensation: To handle the time delay between LiDAR processing and state updates, the system employs an extrapolation strategy. Delayed measurements are projected to the current timestep by exploiting the difference between the predicted state and the state history, allowing for smooth odometry output without large jumps.
• Upsampling: The delay compensation strategy allows for upsampling measurements, distributing their influence over multiple update cycles.

Key Contributions

1. FAR-LIO Framework: Introduction of a highly optimized, CUDA-accelerated LIO method that integrates a sparsity-aware GICP front-end with an EKF backend.
2. Novel Data Structures: Development of a CUDA-based voxel hashmap (cuVoxelMap) with adaptive density, enabling efficient parallelized kNN search and map updates.
3. Ablation and Analysis: Detailed ablation studies of core components (SA-GICP, ASMD, Delay Compensation) and an analysis of the critical relationship between odometry latency and system robustness.
4. Validation: Comprehensive evaluation across four different sensor setups using public datasets (KITTI, MulRan) and real-world data from autonomous racecars (Indy Autonomous Challenge, Abu Dhabi Autonomous Racing League) operating at speeds up to 250 km/h.

Results

The authors evaluated FAR-LIO against state-of-the-art baselines (KISS-ICP, FAST-LIO2, Faster-LIO, D-LIO) using a single parameter set across all environments.

• Runtime Efficiency: FAR-LIO achieved the fastest computation times, outperforming the next-best method by an average of 38.4%. On autonomous racing data (three concatenated LiDARs, ~200k points), the total callback time was 19.23 ms, well below the sensor frequency. In contrast, several CPU-based baselines failed to operate within sensor frequencies or diverged on high-speed data.
• Accuracy: FAR-LIO demonstrated an average 6.9% reduction in positional error (APE) compared to baselines. It achieved the lowest APE and Relative Positional Error (RPE) in the majority of sequences across all datasets. Notably, it was the only method to successfully complete all sequences in the autonomous racing and KITTI datasets without divergence.
• Robustness: The system maintained stability in environments with extreme accelerations and IMU noise. Ablation studies confirmed that the combination of the EKF with delay compensation and the SA-GICP significantly reduced drift and improved accuracy compared to simpler baselines.
• Latency Sensitivity: Experiments simulating additional odometry delays showed that APE remains stable up to a threshold of 60–100 ms, beyond which estimation diverges. FAR-LIO's actual runtime remains well within this safe margin.

Significance and Claims

The paper claims that FAR-LIO successfully bridges the gap between high-accuracy localization and the strict real-time requirements of high-speed autonomy. By leveraging GPU acceleration and novel data structures, the framework demonstrates that it is possible to achieve low-latency odometry without sacrificing accuracy.

The authors emphasize that their approach is specifically designed for real-world deployment, validated on the official compute platform of the Indy Autonomous Challenge (dSpace AUTERA Autobox). The ability to use a single parameter set across diverse environments—from residential areas and highways to racetracks—highlights the system's generalizability. The work underscores that minimizing odometry latency is not merely a performance metric but a critical factor for the stability of closed-loop control in dynamic, high-speed scenarios. The authors have open-sourced the framework to provide a robust baseline for future research in LiDAR-Inertial odometry.