Nomagic details using synthetic sim data to train rare-failure robot anomaly detection
m_wulfmeier · x · 2026-10-28
A Nomagic blog post explains how they trained Live Anomaly Detection (LAD) — a 15 FPS vision system that pauses the robot when an item disassembles mid-pick, escalating to a remote operator — using synthetic data, since real failures are too rare to learn from. Metrics: false negative rate (missed broken items shipped to customers) vs false positive rate (operator time and cycle-time lost). Data is the hard part: positives are 1% prevalence, with an early test set of just 37 positive attempts against 2,604 clean ones, so one extra detection moves true positive rate by 3 points. Their fix: synthesize the rare failure modes in simulation and multiply deployment data instead of waiting for real failures.
More from Embodied
- Dream-1 robot arms grab limes zero-shot from a single prompt in restaurant kitchen demo — ___Mufasaa · 2026-10-28
- Metatron: a System 1 controller that lets coding agents command humanoid robots — ChongZzZhang · 2026-10-28
- DeFM: ETH team builds a DINO-style foundation model for depth images in robotics — ChongZzZhang · 2026-10-28
- Robot fighting teams already run sim2real2sim training pipelines, confirms practitioner cixliv — cixliv · 2026-10-28
- Wayve researcher: there is no good sim2real — it's always real2sim2real — m_wulfmeier · 2026-10-28
- Foldable iPhone retail box surfaces in photo as developer posts "We're doing it" — Dimillian · 2026-10-28