We train industrial perception models on synthetic data — no manual labels, no data collection campaigns, no domain gap.
Five stages, fully automated. Every step is reproducible, every label is generated, and every model improves the next iteration — without a human ever drawing a bounding box.
Photogrammetry on the physical hazard. The asset carries real geometry and real surface response.
Drop into CARLA or NVIDIA Omniverse. The simulator becomes a sampler over every condition the model will face.
All lighting, all angles, all weather. Ground truth — bounding boxes, masks, depth — emitted alongside every frame.
SynYOLO — our YOLOX-based detector — learns directly from synthetic frames. No annotation teams, no labeling queue, no humans in the loop.
TensorRT inference on industrial RTSP feeds. Pixel detections become real-world coordinates via camera calibration.
Real-world data is expensive to collect, dangerous to stage, and impossible to label fast enough. We replace all three problems with a single rendered pixel.
No annotation teams. No data collection campaigns. No labeling queue running ahead of you forever. The simulator emits ground truth — every box, every mask, every depth value — alongside the image it just rendered.
A new hazard class — from 3D capture to production model on a live camera — in days, not quarters. New failure modes in the field become a new render pass overnight. Retrain and redeploy before the next shift.
Models trained entirely in simulation work in reality. Photoreal rendering, domain randomisation, and PBR materials close the gap — verified on live factory footage the model has never seen.
A research-grade smart-manufacturing testbed with permanent industrial cameras. Our prototype detects traffic cones — trained on zero real photographs, deployed against real ones.
Active members of the leading startup programs from NVIDIA, AWS, and Microsoft — and a partner in Austria's 5G LUMEIK testbed.