AnalysisRoboticsSeptember 4, 2026

Hobbyist runs VLM on NVIDIA AGX Thor with RealSense at 10-12 FPS

A robotics hobbyist reports running a BLIP-CUDA vision-language model on an NVIDIA AGX Thor with RealSense 3D stereo vision, achieving 75-100ms per inference (10-12 VLM scene inferences per second). Everything, including the RealSense SDK, runs on GPU.

1 source

More stories today

Open the live feed