nvidia-isaac/cuVSLAM

What it solves

cuVSLAM provides high-performance visual tracking and Simultaneous Localization and Mapping (SLAM) for robots. It solves the problem of achieving real-time, accurate spatial awareness and movement tracking (odometry) without requiring manual parameter tuning, which is often a barrier in complex robotic environments.

How it works

The library leverages CUDA acceleration to process visual data efficiently. It supports multiple tracking modes based on the available sensor hardware:

• Mono: Single-camera tracking (scale-ambiguous).
• RGBD: Single depth-aligned camera.
• Multicamera: Stereo or multi-stereo rigs (up to 32 cameras).
• Inertial: Stereo pair combined with an IMU for increased robustness.
• Multisensor: A flexible mix of RGB, RGB-D cameras, and IMUs (requires cuNLS for nonlinear least squares optimization).

Who it’s for

It is designed for robotics developers and researchers working with NVIDIA hardware (Desktop, Server, or Jetson Orin/Thor) who need a production-ready SLAM solution that works "out of the box" with standard camera calibrations.

Highlights

• CUDA-Accelerated: Optimized for real-time performance on NVIDIA GPUs.
• Zero-Tuning: Uses engineered defaults that adapt automatically to environment and motion profiles.
• Flexible Sensor Support: Handles everything from a single camera to complex multi-camera and IMU-fused rigs.
• Broad API Support: Provides both C++ and Python (PyCuVSLAM) bindings, with integrated ROS2 support via Isaac ROS.