Hardware req

[monajalal]

1. What is minimum GPU required for running cuVSLAM for two Intel RealSense D435i cameras in real-time?
2. Does cuVSLAM work for vertical baselines or only horizontal baselines (no other disparity, two cameras separated vertically 230mm) on a fixed 3D printed mount on a moving cart (I move the cart slowly)?
3. Does HW sync work out of box for these two cameras and Orin NX 16G and does cuVSLAM work in real-time in Orin NX 16G?

I saw your tutorial was designed for D455 and Thor so just wanted to get some confirmation before I invest my time heavily on these.

multi realsense cuvslam on Thor https://nvidia-isaac-ros.github.io/concepts/visual_slam/cuvslam/tutorial_multi_realsense.html

multi realsense hardware sync on Thor https://nvidia-isaac-ros.github.io/robots/sensor_mounting_rig/thor_devkit_realsense_rig_assembly.html
Thank you.

[zwdoescode]

1. It depends on image resolution and tracking mode. For VGA even Jetson Orin Nano should be OK, depend on how much GPU/CPU load you are ready to spend on odometry.   Details here: https://arxiv.org/pdf/2506.04359 System Performance and Hardware Utilization
2. Yes it is. User only need up/down view overlap and proper calibration.
3. Let me get more information from our developers.

[aefitorov-nvidia]

Hi, @monajalal, thank you for your interest in multi-camera RealSense support.

Q1: Minimum GPU Requirements
A Jetson Nano is sufficient: As detailed in our technical report, the Jetson Nano can successfully handle 3 RealSense D435 cameras with cuVSLAM live tracking odometry running in real-time (not just replay).

Yes, the Orin NX 16G is more than capable of running cuVSLAM in real-time, given that the less powerful Jetson Nano can handle the workload.

Recommendation: use 640x480 at 60fps profile for the best experience. Increasing resolution beyond this does not significantly improve quality, however, maintaining 60fps is beneficial for handling fast robot turns and movements.

Q3: HW Sync and Orin NX Support

The D435i cameras connected via USB support Hardware Synchronization, but it requires specific setup (please read Multi-Camera RealSense Hardware Assembly Guide for Jetson Orin):

Critical: A physical wired connection via JST connectors between cameras (since you are using 2 cameras a direct wire connection could be used)
Critical: Configuring cameras to "Master-Slave" mode during initialization with proper parameters.
Recommended: Building librealsense from source on the device with the specific flags enabled.

Recommendation: try the multi-RealSense cuVSLAM via Python bindings first: setting up camera parameters and rigs is generally easier in Python for initial testing

I will be happy to help if other questions arise

[monajalal]

@aefitorov-nvidia @zwdoescode could you please share what is the exact license of this? I understand this is NVIDIA Community license? Does it mean I can use it at our company Toyota Material Handling North America with no problem? We want to clear out the licensing before diving in further. Thanks a lot for your detailed response.

I am confused by the response I got from Enterprise ChatGPT

[zwdoescode]

@monajalal The license allows anyone to freely use cuVSLAM and PyCuVSLAM for commercial purposes given the system which includes cuVSLAM and PyCuVSLAM run on NVIDIA hardware.
So yes, you can use it at TMH if your vehicle includes a Jetson.
Your GPT answer about PyCuVSLAM is based on our older license. But now cuVSLAM and PyCuVSLAM are using the same license.