ZereData generates custom synthetic training data for robotic bin picking.

Bin picking is one of the hardest vision problems in warehouse automation. Training models requires 6D pose ground truth, rotation plus translation for every object in a cluttered bin, which cannot be cost-effectively collected from real-world operations. Manual annotation is impossible at scale. Real-world data collection runs $2–5 per labeled image. Academic datasets don’t match any real warehouse’s SKU catalog.

We render photorealistic bin-picking scenes from customer CAD files or reference photos, with physics-based ray tracing and domain randomization. Every scene includes 6D pose for every object, instance segmentation masks, depth maps, and bounding boxes, exported in BOP format (the 6D pose estimation standard) along with COCO and YOLO. It’s the data layer that warehouse automation and physical AI depend on.

Pre-seed, based in Rotterdam. The engine is production-ready, with a public sample dataset on Hugging Face. We’re now running our first dataset assessment with a warehouse automation team, generating samples from their own SKUs.

We’re looking for more bin-picking teams open to a short discovery conversation about how they currently approach training data. No pitch, no demo, a learning conversation. We’re also open to design partners interested in generating custom datasets for their specific SKUs.

Reach us at engineering@zeredata.com.

ZereData

Org chart

Teams

Offices

ZereData

Org chart

Teams

Offices

Related companies