Nomagic details simulation pipeline that halves robot anomaly misses at 1% FPR
m_wulfmeier · x · 2026-10-20
Nomagic senior ML research engineer Szymon Trochimiak details how the company uses simulation to multiply deployment data for physical AI, training its Live Anomaly Detection (LAD) system — a vision model that watches every pick at 15fps and pauses the robot after 3 consecutive positive frames for remote operator review. A miss ships medication with missing content; a false alert costs operator attention and cycle time.
Why data is hard:
- Positives are rare (1% prevalence): the held-out test set had 37 positive attempts vs 2,604 clean ones, so one extra detection moves TPR by almost 3 points
- Video is plentiful but useful labels are not
Approach: a synthetic pipeline that extracts item cutouts, matches and overlays them onto real frames, adjusts box position, and applies gamma correction to synthesize anomalies online during training — exposing LAD to failures production rarely shows.
Results: FNR (missed disassembled boxes) reduced 2x at only 1% FPR. Takeaway: simulation and real-world data are complementary, not competing.
More from Embodied
- Eka Robotics raises $150M Series A at $1.45B valuation for its Vision-Force-Action manipulation model — pulkitology · 2026-10-20
- Arda exits stealth with AI agent that turns CAD files into assembly instructions, deployed across Anduril — pxd · 2026-10-20
- Mac Studio with 256GB RAM now shipping in 15-17 weeks as demand outstrips supply — LinusEkenstam · 2026-10-20
- RealSense podcast demos Perception Studio: VIO, 2cm min-depth, people detection — chrismatthieu · 2026-10-20
- Mirror learns Minecraft skills from video plans with no rewards or labels in ~30 minutes — vincesitzmann · 2026-10-20
- Tinkering a discontinued Facebook Portal into a Meta Muse smart display — msg · 2026-10-20