Nomagic uses simulation to cut robot anomaly misses 2x at 1% false positive rate
m_wulfmeier · x · 2026-10-20
Robotics company Nomagic shared at NVIDIA GTC Berlin how it complements real deployment data with simulation to train its Live Anomaly Detection (LAD) system. LAD watches every pick at 15fps, pausing the robot after 3 consecutive positive frames for human review. With anomalies at 1% prevalence (37 positives vs 2,604 clean picks in the test set), they built a synthetic data pipeline (cutout extraction, matching, overlay, repositioning, gamma correction) to teach LAD rare failure modes — halving FNR at only 1% FPR.
Related event: Nomagic Uses Simulation to Halve Anomaly Miss Rate in Physical AI(2 posts)→
More from Embodied
- Wayve takes its autonomous Fiat 500e up Turin's iconic Lingotto rooftop track — Sethwinterroth · 2026-10-20
- Waymo Logs 95% Fewer Serious Crashes Than Humans, Eyes 20,000-Car Fleet by 2027 — Anen-o-me · 2026-10-20
- a16z leads Series A in Eka Robotics, betting on force-aware dexterous manipulation — a16z Newsletter · 2026-10-20
- Eka Robotics hits $1.45B valuation as experts rave over its object-picking arm — TechCrunch AI · 2026-10-20
- Agility founder: "I cannot love Digit 5" — SPAC merger to take robot maker to Nasdaq — MarwaEldiwiny · 2026-10-20
- Sepp Hochreiter's NXAI partners with NEURA Robotics to build memory for cognitive robots — HochreiterSepp · 2026-10-20