NVIDIA AGX Thor + RealSense hits 75ms per VLM depth inference, ~10-12 FPS scenes

chrismatthieu · x · 2026-09-04

Day 2 of a robotics hack on NVIDIA's AGX Thor with RealSense 3D stereo vision: everything including the RealSense SDK now runs on GPU, the author swapped Qwen for BLIP-CUDA so the robot describes its surroundings with depth at 75ms per inference (10-12 VLM scene inferences per second). Source code shared in comments.

Related event: Open-Sourced VLM System Gives Robots Real-Time Scene Understanding on Jetson AGX Thor(3 posts)→

Original post →

More from Embodied

Embodied channel →