Breaking

NVIDIA Open-Sources cuVSLAM, a CUDA-Accelerated SLAM Library

June 28, 2026 at 10:34 EDT

Robotics
Open Source
Infra & Chips

NVIDIA has open-sourced cuVSLAM on GitHub, a library for CUDA-accelerated visual odometry and mapping released as part of its Isaac project. It supports up to 32 cameras plus an IMU and is designed for real-time operation on Jetson-class edge devices.

NVIDIA Isaac · cuVSLAM Release

GPU-Accelerated Visual SLAM Brings Real-Time Localization to the Edge

A CUDA-optimized visual odometry & mapping library — C++ core, Python bindings, and ROS 2 — built to run with low overhead on Jetson, supporting up to 32 cameras plus IMU and depth.

cameras max, arbitrary geometry

<1%

avg. trajectory error on KITTI

<5cm

position error on EuRoC

sensor modes supported

Front-end / Back-end split

Low-latency pose estimation runs up front while map refinement happens asynchronously.

Front-end

Low-latency visual odometry & pose estimation

→

Back-end

Async map refinement, loop closure, PGO & CUDA bundle adjustment

Multi-Stereo vs Single Stereo

Scaling from one stereo pair toward 32 freely-arranged cameras raises accuracy and robustness.

Single Stereo

2 cameras

→

Multi-camera

up to 32 cameras

Developer praise

High accuracy on EuRoC & KITTI
ROS 2 integration practical for navigation
Easy live-camera / dataset feed via Python API
Low CPU/GPU load on Jetson Orin AGX

Open challenges

CUDA 13.1 warnings & limitations
IMU input drop behavior
Unstable on lower-spec Orin Nano (AGX advised)
Sensitive to calibration, sync & motion blur

Positioned as a localization foundation across the NVIDIA Isaac ecosystem.

C++ library + PyCuVSLAM + ROS 2 Ubuntu 22.04 / 24.04 CUDA 12 / 13 x86_64 & aarch64 (Jetson) Ties into Perceptor & nvblox

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.