Real-Time Robotics: Lightweight, Sovereign AI | Clément Busuttil, Neuronaix | FUSE Toulouse 2026

A robot arm controlled at 500 Hz can’t stop and wait for a neural network. Clément Busuttil, President of Neuronaix, shows how his team put a standard six-axis robot on an aerospace production line to insert hair-thin wires for wire-wrapped electronics, a task that takes human operators well over a year to master. He explains how they cut inference from about 200 ms to camera frame rate by seeding with a large model and then tracking a small region of interest, fused stereo vision with the robot’s joint data to reconstruct the wire in 3D and compensate for joint drift, and built lightweight models for an embedded GPU, trained on synthetic data and refined with real-world failure cases, all running on-site with no data leaving the plant. Topics: real-time control vs inference latency, region-of-interest perception, vision and joint-state fusion, automated quality checks, robots with no pose feedback, synthetic training and adversarial cases, and keeping perception and control separate so the pipeline can be audited. Recorded at FUSE Toulouse, October 6, 2026, Orange Innovation Center, Balma, France. FUSE is a global technical conference series for localization and Physical AI engineers. Next: FUSE San Francisco, November 19, Foundry SF. Register: https://fuse.pointonenav.com/sf Chapters 00:00 The reverse problem: localizing the workpiece, not the robot 00:46 Real-time constraints: 500 Hz control vs model inference 02:17 The pipeline: cameras, robot state, perception, control 03:27 From 200 ms inference to camera frame rate 06:23 Hybrid approach: seed with a big model, track a small region 07:02 Use case: wire-wrapping for helicopter electronics 10:00 Fusing stereo vision with joint data to reconstruct the wire 12:41 Automated quality control: detecting buckling 16:23 Nuclear: robots with no pose feedback 18:40 Lightweight models, synthetic training and adversarial cases 22:12 Turning tradesman knowledge into the stack 24:00 Q&A: inputs, modular vs end-to-end, security, success rates