Lidar-inertial odometry (LIO) has an efficiency problem. High-resolution lidars produce millions of points per second that must be processed online on the resource-constrained hardware of robots, cars, or drones, yet most LIO methods neither bound their memory nor report their compute footprint. We introduce Pelorus, a resource-aware LIO with a fixed-capacity multiresolution map of voxelized normal distributions and an iterated equivariant state estimator. A relevance-based eviction keeps the map memory bounded, and a maturity-weighted map update suppresses inconsistent observations. The pipeline runs in parallel on CPUs and GPUs. On five sequences from three public datasets, Pelorus achieves the lowest relative pose error in seven of ten metrics with a small, bounded memory footprint and, with GPU offloading, the lowest CPU load of all compared methods. On an NVIDIA Jetson Orin NX, it processes a 128-beam lidar at 10 Hz. Pelorus is open source with CUDA, Vulkan, and Metal backends.
Figures & tables
Fig. 1: Memory and logical CPU cores of LIO systems on helipr/town02:ouster in logarithmic scale, measured on an Intel Core i7-14700 with an NVIDIA GeForce RTX 4060 Ti and bounded by its 28 logical cores and 64GB . Boxes span from the 50th percentile to the maximum with the measurement density inside, lines mark the 95th percentiles, and the box color encodes the relative pose error (RPE) as a multiple of the lowest one on a logarithmic scale. The RPE in the labels is the geometric mean of its translational and rotational parts. Pelorus achieves the lowest relative error with a small, bounded memory footprint. Only RKO-LIO uses less memory, but at a higher RPE and CPU load. With CUDA, Pelorus has the lowest CPU load.
Dataset
Metric
Pelorus
FAST-LIO2 [ 6 ]
Eq-LIO [ 13 ]
Faster-LIO [ 9 ]
RKO-LIO [ 36 ]
EllipseLIO [ 37 ]
DLIO [ 10 ]
iG-LIO [ 19 ]
KISS-ICP [ 38 ]
Newer College [ 29 ]
RPE T
0.0255
0.0373
0.0735
0.0310
0.0496
0.0498
0.0529
0.0276
0.2468
park_merged:os0_128
RPE R
0.2821
0.4795
0.5227
0.4815
0.4281
0.4762
1.3187
0.2913
0.8397
2.4km
APE T
0.2925
0.2655
11.4013
0.2134
1.4669
0.2970
0.2749
0.1973
0.7702
APE R
0.6632
0.6196
5.8040
0.5851
1.0732
0.6357
1.3189
0.4782
1.0919
HeLiPR [ 30 ]
RPE T
0.0648
0.0611
d.n.f.
0.0628
0.0907
d.n.f.
0.0865
0.2487
0.1302
bridge02:ouster
RPE R
0.0568
0.0536
d.n.f.
0.0437
0.0582
d.n.f.
0.0826
0.1230
0.0689
TABLE I: Accuracy Across LIO Systems
Method
Scans
Latency
Proc.
Power
E ⋅ RPE
(%)
p95 (ms)
p95 (ms)
(W)
(mJ m)
Pelorus Vulkan
100.0
192
97.8
13.1
50.3
Pelorus CUDA
99.7
175
83.7
12.4
48.4
Pelorus CUDA + CPU ×4
100.0
173
78.1
12.2
46.8
Pelorus CPU ×4
87.6
3105
220.3
12.0
50.7
Pelorus CPU ×8
93.1
2616
173.2
12.4
49.9
TABLE II: Timings, Power, and Energy Efficiency on the Jetson Orin NX
Fig. 2: Resource footprints on an NVIDIA Jetson Orin NX running helipr/town02:ouster in real time, with the axes bounded by its eight CPU cores and 16GB . With CUDA, Pelorus handles the high-bandwidth OS2 lidar at sensor rate, and both builds hold the desktop accuracy with a fraction of the resources of the other methods. FAST-LIO2, Faster-LIO, and KISS-ICP drop too many scans and lose track, and EllipseLIO and Eq-LIO crash on the board, so they are not shown. Table II lists all methods.
Fig. 3: Capacity sweep on helipr/town02:ouster with the CPU build. Translational and rotational APE and RPE over the peak resident memory for map capacities ∣M∣ from 32k to 512k distributions (k = 1024), each relative to the smallest capacity. The capacity used is marked in bold.
APE T (m)
RPE T
∣M∣
Eviction
NC
OQ02
CC02
town02
( % )
16384
LRU
5.36
0.728
0.411
16.0
T/P
2.15
0.759
0.226
18.2
+2
32768
LRU
1.88
0.404
0.187
9.16
T/P
1.32
0.181
0.132
9.75
+1
65536
LRU
2.10
0.365
0.217
6.01
TABLE III: Eviction Ablation with Reduced Map Capacity
APE T (m) and change ( % )
Mean ( % )
Update
NC
OQ02
CC02
APE T
RPE T
Count-weighted
0.309
0.224
0.125
0
0
κ=5
0.308 (+0)
0.123 ( − 45)
0.101 ( − 19)
− 22
+5
κ=10
0.292 ( − 5)
0.097 ( − 57)
0.089 ( − 29)
− 30
+5
κ=20
0.317 (+3)
0.120 ( − 46)
0.102 ( − 19)
− 21
+7
κ=40
0.298 ( − 3)
0.145 ( − 35)
0.109 ( − 13)
− 17
+10
TABLE IV: Map Update Ablation
RPE T (m)
Map
NC
OQ02
CC02
town02
Multiresolution
0.0255
0.0073
0.0071
0.0391
Single 0.3m
0.0261
0.0077
0.0087
0.0404
Single 0.5m
0.0284
0.0127
0.0153
0.0438
Single 0.7m
0.0330
0.0211
0.0383
0.0492
TABLE V: Multiresolution Ablation with Equal Total Capacity
LiDAR Inertial Odometry (LIO) is a critical component for many mobile robots that need to navigate without relying on external positioning (e.g., GPS). Platforms that operate autonomously in different environments and with heterogeneous LiDAR sensors require a LIO approach that can adapt to these different scenarios without human intervention. Existing LIO approaches can typically provide reliable and accurate odometry in scenarios with similar environments and sensors when suitably tuned. However, many approaches struggle to retain robust odometry across heterogeneous environments and sensors while using a consistent configuration. This paper presents EllipseLIO, a real-time LIO approach that generalises between scenarios by using methods for LiDAR scan filtering and registration that adapt to the sensor capabilities and environment without requiring scenario-specific tuning. Experiments with EllipseLIO and state-of-the-art LIO approaches on five datasets with diverse and challenging scenarios demonstrate that EllipseLIO is the best-performing approach overall. It achieves a 38% lower odometry error on average than the second-best approach and is the only approach that does not diverge in any experiment. An open-source version of EllipseLIO will be available at github.com/v4rl-ucy/ellipselio.
Rowan Border, Margarita Chli
Vision for Robotics Lab (V4RL), University of Cyprus, Cyprus · ETH Zurich, Switzerland
Robust and accurate odometry estimation is essential in modern robotics. In environments characterized by highly dynamic motion and sensor noise, odometry estimation becomes increasingly challenging. Autonomous racing combines both factors in an unstructured setting, where minimizing odometry latency is essential for stable closed-loop control. This paper introduces FAR-LIO, a highly optimized CUDA-accelerated LiDAR-inertial odometry framework developed for Fast, Accurate, and Robust performance. Our system leverages a novel CUDA-based voxel hashmap to enable parallelized nearest-neighbor search and efficient map updates. We employ a sparsity-aware Generalized Iterative Closest Point algorithm with adaptive thresholding on top of the CUDA-based voxel hashmap with adaptive density to achieve low-latency without compromising accuracy. An Extended Kalman Filter serves as a robust backend. It utilizes an upsampling and delay compensation strategy to fuse the LiDAR odometry with high-frequency IMU data, thereby ensuring a robust and smooth odometry output. We evaluate FAR-LIO across four different sensor setups, using both public datasets and data from two autonomous racecars driving at speeds of up to 250 km/h. FAR-LIO achieves an average 6.9% reduction in the positional error and 38.4% lower runtime compared to state-of-the-art baselines on target hardware using a single parameter set. This demonstrates its computational efficiency and broad applicability. To build upon our work, our code is available open-source on https://github.com/TUMFTM/FAR-LIO.
Maximilian Leitenstern, Marcel Weinmann, Patrick Haft +3
Institute of Automotive Technology, School of Engineering and Design, Technical University of Munich, Germany · NVIDIA Corp., USA
Measurement-level LiDAR-Inertial-Visual Odometry (LIVO) can achieve high accuracy, and its practical resource requirements depend on the adopted association and map-maintenance strategy. Persistently attaching dense photometric patches or appearance attributes to mapped LiDAR points can introduce additional storage, association, and update costs. This paper presents F-LIVO, a filter-based LIVO framework that couples a visual-inertial subsystem and a LiDAR-inertial subsystem through feature-level point-to-plane constraints between visual points and LiDAR planes. To preserve cross-modal coupling when visual triangulation is weak, we further recover visual feature depth from the LiDAR map through local plane fitting and ray-plane intersection. In the LiDAR update, the covariance of each point-to-plane residual is adaptively scaled according to local planar reliability, thereby reducing overconfident updates on poorly conditioned surfaces. Because the proposed method relies on sparse geometric features rather than persistent dense photometric map attributes, it retains a compact map representation. Experiments on multiple public datasets show competitive, sequence-dependent accuracy. Across the evaluated degradation cases, F-LIVO maintains continuous estimation with comparatively low accumulated end-to-end drift. Under the evaluated resource configuration, it achieves the lowest mean and peak central processing unit (CPU) usage, the lowest peak resident set size (RSS), and the lowest RSS growth rate among the evaluated implementations, while R2LIVE retains a lower mean RSS. The source code will be made publicly available upon acceptance to facilitate reproducible research.
Xinye Dai, Yapeng Wang, Shujie Zhou +6
The authors are with the GNSS Center, Wuhan University, Wuhan 430079, China.