Industrial cuVSLAM Benchmark & Integration

Charbel Abi Hana; Kameel Amareen; Mohamad Mostafa; Dmitry Slepichev; Hesam Rabeti; Zheng Wang; Mihir Acharya; Anthony Rizk

cs.ROMar 17, 2026

Industrial cuVSLAM Benchmark & Integration

Authors: Charbel Abi Hana, Kameel Amareen, Mohamad Mostafa, Dmitry Slepichev, Hesam Rabeti, Zheng Wang, Mihir Acharya, Anthony Rizk

Organizations: Idealworks · NVIDIA

Abstract

This work presents a comprehensive benchmark evaluation of visual odometry (VO) and visual SLAM (VSLAM) systems for mobile robot navigation in real-world logistical environments. We compare multiple visual odometry approaches across controlled trajectories covering translational, rotational, and mixed motion patterns, as well as a large-scale production facility dataset spanning approximately 1.7 km. Performance is evaluated using Absolute Pose Error (APE) against ground truth from a Vicon motion capture system and a LiDAR-based SLAM reference. Our results show that a hybrid stack combining the cuVSLAM front-end with a custom SLAM back-end achieves the strongest mapping accuracy, motivating a deeper integration of cuVSLAM as the core VO component in our robotics stack. We further validate this integration by deploying and testing the cuVSLAM-based VO stack on an NVIDIA Jetson platform.

Figures & tables

Figure 1: VO trajectories (from left to right): L-shape, hybrid & rotation-heavy sequences.

Sequence	Linear Velocity (m/s)		Angular
Sequence	Mean	Max	Mean Angular Vel. (deg/s)	Max Angular Acceleration (deg/s 2 )
L-shaped	0.53	2.05	6.44	16.69
Rotation-Heavy	0.31	1.00	29.33	37.80
Hybrid	0.42	2.76	25.35	39.28

Table 1: Trajectory motion statistics per VO sequence.

Method	L-shaped	Rotation-Heavy	Long Sequence	Avg. Inference	Max Inference	Avg. CPU	Max CPU	AVG GPU
				Time (ms)	Time (ms)	Usage (%)	Usage (%)	Usage (%)
cuVSLAM	$\mathbf{0.057\pm 0.01}$	$\mathbf{0.296\pm 0.025}$	$\mathbf{0.850\pm 0.034}$	$\mathbf{2.262}$	$\mathbf{3.137}$	$\mathbf{2.002}$	$8.600$	$6.200$
ORB-SLAM	$0.103$	$0.392$	$0.974$	$21.828$	$30.074$	$4.759$	$11.500$	$-$
RTAB-Map	$0.10\pm 0.02$	$0.390\pm 0.14$	$1.37\pm 0.406$	$35.033$	$45.354$	$3.573$	$\mathbf{7.200}$	$-$
Wheel Encoder	$0.071$	$0.183$	$0.604$	$-$	$-$	$-$	$-$	$-$

Table 2: APE (m) & Computational Costs results across scenarios. APE: Mean

\pm

std over 10 runs (single run for ORB-SLAM and Wheel Encoder) on Desktop

Method	L-shaped	Rotation-Heavy	Long Sequence	Avg. Inference	Max Inference	Avg. CPU	Max CPU	AVG GPU
				Time (ms)	Time (ms)	Usage (%)	Usage (%)	Usage (%)
cuVSLAM	$\mathbf{0.067\pm 0.05}$	$\mathbf{0.301\pm 0.12}$	$\mathbf{0.640\pm 0.43}$	$\mathbf{9.400}$	$\mathbf{40.340}$	$\mathbf{9.601}$	$61.011$	$5.300$
ORB-SLAM	$0.150$	$0.308$	$0.899$	$60.342$	$71.572$	$36.009$	$61.500$	$-$
RTAB-Map	$0.11\pm 0.06$	$0.311\pm 0.17$	$0.92\pm 0.54$	$86.300$	$144.7$	$25.600$	$\mathbf{40.500}$	$-$
Wheel Encoder	$0.071$	$0.183$	$0.604$	$-$	$-$	$-$	$-$	$-$

Table 3: APE (m) & Computational Costs results across scenarios. APE: Mean

\pm

std over 10 runs (single run for ORB-SLAM and Wheel Encoder) on Xavier

Figure 2: Radar plot comparing the metrics of each method across the three VO sequences. The wheel encoder baseline is included for reference. Smaller areas indicate better performance.

Method	Seq-570			Seq-1700			Tick-Rate (Hz)	Avg. CPU Usage (%)	Max CPU Usage (%)	Avg. GPU Usage (%)
Method	Mean	Max	RMSE	Mean	Max	RMSE	Tick-Rate (Hz)	Avg. CPU Usage (%)	Max CPU Usage (%)	Avg. GPU Usage (%)
cuVSLAM	$\mathbf{0.32\pm 0.25}$	$1.27$	$0.41$	$1.57\pm 1.25$	$5.21$	$2.01$	$\mathbf{105.3}$	$\mathbf{9.10}$	$\mathbf{33.00}$	$\mathbf{65.00}$
Idealworks NVIDIA Hybrid	$0.61\pm 0.29$	$1.09$	$0.68$	$\mathbf{0.91\pm 0.29}$	$\mathbf{2.2}$	$\mathbf{0.54}$	$1.50$	$18.74$	$35.00$	$66.26$
RTAB-Map	$-$	$-$	$-$	$5.89\pm 4.69$	$22.14$	$7.53$	$0.95$	$-$	$-$	$-$
ORB-SLAM	$0.34\pm 0.22$	$1.61$	$0.38$	$1.17\pm 1.16$	$5.19$	$1.65$	$12.50$	$52.28$	$79.00$	$-$

Table 4: VSLAM APE (m) results on the deployment dataset on Xavier

Method	Large-Scale Support	Inference Time	Production Grade	Rank
Idealworks NVIDIA Hybrid	$\oplus$ Post-run map management $\oplus$ Multi-session & multi-robot mapping support	$\oplus$ GPU-accelerated pipeline powered by cuVSLAM $\oplus$ Consistent inference time with growing map size	$\oplus$ Production-ready $\oplus$ Seamless ROS2 integration	1 st
cuVSLAM	$\oplus$ Multi-camera support $\oplus$ Multi-modality support (IMU, RGB-D, Stereo) $\ominus$ No post-run map management tools	$\oplus$ GPU-accelerated $\oplus$ High throughput from SLAM stack	$\oplus$ Production-ready $\oplus$ ROS2 integration via Isaac ROS Visual SLAM	2 nd
RTAB-Map	$\oplus$ Memory management system $\oplus$ Constant inference time on unbounded maps $\oplus$ Multi-robot & multi-session support	$\ominus$ Dense stack, generally slower $\ominus$ Unstable odometry under rapid motion $\sim$ Inference time stable with map size	$\oplus$ Near production-ready $\sim$ Moderate integration effort	3 rd
ORB-SLAM3	$\ominus$ No large-map handling $\ominus$ No pure localization mode $\ominus$ No multi-session or multi-robot support	$\oplus$ Fast map updates via multi-threading $\ominus$ Slow ORB feature extraction $\ominus$ Inference time degrades with map size	$\oplus$ Competitive accuracy in industrial settings $\ominus$ Heavy development effort required $\ominus$ Limited production support	4 th

Table 5: Deployability comparison of VSLAM methods.

Panel of Figure 3 — Figure 3: Idealworks NVIDIA hybrid estimated trajectory errors across both sequences.

Visual Odometry

Visual odometry estimates camera or robot motion from successive images, enabling navigation and mapping when external positioning is unavailable.

See all 56 papers

Line-Augmented cuVSLAM: Robust Visual Odometry in Low-Texture Environments
Oct 14, 2026Dara Molloy, Terry Wu, Wonju Lee +4Visual Odometry Simultaneous Localization and Mapping
∗Analog Devices · †NVIDIA
Smart Fusion for Adaptive Visual-IMU Weighting in cuVSLAM Sparse Bundle Adjustment
Oct 14, 2026Wonju Lee, Dara Molloy, Abhilash Nand Kumar +5Visual-Inertial SLAM Visual-Inertial Odometry
Analog Devices Inc. · NVIDIA
Robust Visual SLAM for UAV Navigation in GPS-Denied and Degraded Environments: A Multi-Paradigm Evaluation and Deployment Study
May 5, 2026Prasoon Kumar, Akshay Deepak, Sandeep KumarUAV Navigation GNSS-Denied Localization
National Institute of Technology Patna, Bihar, India · Central Research Laboratory, Bharat Electronics Ltd., Ghaziabad, Uttar Pradesh, India

Oct 14, 2026J/K move · Enter open · S save

Line-Augmented cuVSLAM: Robust Visual Odometry in Low-Texture Environments

Dara Molloy, Terry Wu, Wonju Lee, Abhilash Nand Kumar+3

∗Analog Devices · †NVIDIA

Visual Odometry Simultaneous Localization and Mapping

Open paper AskPDF arXiv

Autonomous mobile robots in warehouses and production facilities operate in precisely the environments that defeat point-feature visual odometry: featureless aisles, uniform walls, repetitive racking and specular floors. Such environments are, however, rich in straight-edge structure: rack uprights, door frames, wall junctions and cabinet outlines present high-contrast linear edges stable across wide viewpoint changes. A line-feature extension to cuVSLAM, a CUDA-accelerated visual-SLAM system, is presented that recovers this signal. Lines are parameterised as Plücker coordinates for geometric transport and as the minimal 4-DOF orthonormal representation of Bartoli and Sturm for optimisation; a Cauchy-robust point-to-line residual with analytic Jacobians is folded directly into cuVSLAM's per-frame Gauss-Newton pose solve, and the same line state is jointly refined in the Schur-complement sliding-window bundle adjustment. An adaptive information weight causes lines to assume more of the estimation burden exactly as point features become scarce, and the extension adds residual blocks to existing solvers rather than introducing a second estimator, reducing exactly to the unmodified point pipeline when disabled. Evaluated in RGB-D mode under SE(3) Umeyama alignment with no scale correction, line features reduce mean absolute trajectory error by 38.5% over the 22 sequences of the TUM RGB-D Freiburg3 collection and by 30.1% over the 8 Replica room and office sequences, where median translation error falls by 38.9% and median rotation error by 66.4%, with zero tracking failures in any configuration tested.

Abstract

Visual Odometry

Explore similar work

Line-Augmented cuVSLAM: Robust Visual Odometry in Low-Texture Environments

Smart Fusion for Adaptive Visual-IMU Weighting in cuVSLAM Sparse Bundle Adjustment

Robust Visual SLAM for UAV Navigation in GPS-Denied and Degraded Environments: A Multi-Paradigm Evaluation and Deployment Study