// the find
NVIDIA-ISAAC-ROS/isaac_ros_visual_slam
Visual SLAM/odometry package based on NVIDIA-accelerated cuVSLAM
GPU-accelerated stereo visual SLAM from NVIDIA, wrapping their proprietary cuVSLAM library as a ROS 2 node. Targets Jetson hardware for real-time robot navigation without GPS. Best fit for robotics engineers building autonomous systems on NVIDIA's embedded platform.
Benchmarks against ORB-SLAM2 on KITTI show roughly 8x faster runtime (7ms vs 60ms) with better accuracy — that gap is real and matters for real-time nav. Multi-camera support with configurable YAML per sensor (ZED, RealSense, Hawk) means you're not locked to one stereo rig. Map save/load and localize-in-map services let you pre-build a map once and relocalize later, which is the actual production SLAM workflow. The message stream sequencer in the impl layer handles temporal sync between camera frames and IMU data, which is the part most DIY SLAM integrations get wrong.
The SLAM core is a closed binary (cuVSLAM) — you cannot inspect, modify, or build the algorithm itself; you're just wrapping NVIDIA's precompiled library. This is Jetson-or-nothing in practice; the benchmark table only lists aarch64 and NVIDIA's own Nova Orin system, so x86 dev machines need a workaround or Isaac Sim. Python tests are POL (proof-of-life) style — they launch the node and check it doesn't crash, not that the odometry is actually accurate, so regressions in tracking quality won't be caught automatically. No fallback or graceful degradation path documented for feature-poor environments; the README mentions IMU takeover but there's no data on drift rates when running VIO-only for extended periods.