NVIDIA Opens cuVSLAM, Its GPU-Powered Robot Navigation Stack Running 0.4ms per Frame
NVIDIA open-sources cuVSLAM, a CUDA-accelerated visual SLAM library supporting 1–32 cameras with real-time performance on Jetson edge hardware
Takeaways
- NVIDIA open-sourced cuVSLAM, a CUDA-accelerated visual SLAM library with Python, C++, and ROS2 APIs.
- Supports 1 to 32 cameras in arbitrary configurations, plus optional IMU and depth sensors, across 5 tracking modes.
- Achieves sub-1% trajectory error on KITTI and under 5cm position error on EuRoC, running in real-time on Jetson Orin AGX.
- Stereo tracking takes 0.4ms on RTX 4090 and 1.8ms on Jetson AGX Orin — well within 30 FPS budgets.
- Multi-camera mode survives random camera occlusions and delivers ~40% better SLAM accuracy than single stereo on warehouse data.
- Pre-built wheels available for Ubuntu 22.04/24.04, CUDA 12/13, x86_64 and aarch64; NVIDIA-GPU-only dependency.
NVIDIA has open-sourced cuVSLAM, a production-grade visual odometry and SLAM library that runs entirely on the GPU. It ships with a Python API, a C++ API, and full ROS2 integration, and it's already deployed inside NVIDIA's Isaac Perceptor robotics stack. With 1.6k GitHub stars and a freshly published technical paper, this is the clearest look yet at what NVIDIA has been running on its robots.
What problem does it solve?
Visual SLAM , knowing where you are in the world using only cameras , is a foundational capability for any autonomous robot. The classic tension is between accuracy and speed: existing systems either run fast but drift, or run accurately but require a workstation. cuVSLAM directly addresses real-time performance on resource-constrained platforms, limited sensor flexibility, and hardware underutilization. The CUDA-first design means the heavy math , feature tracking, bundle adjustment, loop closure , all happens on the GPU, freeing the CPU for the rest of the robot stack.
A surprisingly flexible sensor model
cuVSLAM can operate with a variety of visual-inertial sensor suites, including multiple RGB and depth cameras and IMUs, supporting configurations from as few as one RGB camera to as many as 32 cameras in arbitrary geometric arrangements. That flexibility is rare. Most SLAM systems are hardcoded to a specific rig.
The library ships five distinct tracking modes, each targeting a different hardware setup:
- Mono , single RGB camera, scale-ambiguous but lightweight
- Mono-Depth , RGB-D camera (e.g. RealSense D435), dense frame-to-frame alignment
- Stereo , the default mode, metric scale from a stereo pair
- Stereo-Inertial , stereo + IMU fusion via a factor graph, robust to fast motion