Give the robot a reliable picture of the cell

Cameras, calibration and models that tell the arm where each part is and which way it faces: in clutter, under changing light, and on parts that shine or are see-through.

Start a project How we choose models

Published accuracy of a Zivid 2+ structured-light camera at 1 m. A wrist RealSense D405 is ±2% at 50 cm.: ±0.2 mmPublished accuracy of a Zivid 2+ structured-light camera at 1 m. A wrist RealSense D405 is ±2% at 50 cm.
The licence of Ultralytics YOLO. For client products we default to permissive detectors such as RF-DETR.: AGPL-3.0The licence of Ultralytics YOLO. For client products we default to permissive detectors such as RF-DETR.
The licence of FoundationPose. When you intend to ship, we use BSD-licensed HappyPose / MegaPose instead.: Non-commercialThe licence of FoundationPose. When you intend to ship, we use BSD-licensed HappyPose / MegaPose instead.

What we do

Robot perception and vision
From choosing the camera to the model that reads it.
- Depth cameras chosen on measured accuracy
- Hand-eye calibration
- Detection and segmentation with RF-DETR, YOLO and SAM 2
- 6-DoF pose estimation for known and new objects
- Point clouds for bins, shelves and trays
- Transparent and reflective parts

Model or no model

Where is the part, and which way is it facing?

Programmed approach

Fixed camera poses and fiducial markers (AprilTag, ArUco)
Color and depth thresholds
Template matching against CAD models

Learned approach

Detection and segmentation models (RF-DETR, YOLO, SAM 2)
6-DoF pose estimation
Trained on synthetic images plus a small real set

How we choose

If parts, fixtures and lighting are controlled, classical vision is cheaper and easier to certify. With many part types, clutter, or shiny and transparent parts, learned perception usually wins. Borrowed weights rarely work as downloaded, so we fine-tune on your cell and prove it on a held-out test set.

Choosing a depth camera

Published accuracy, read from each maker’s datasheet. The right camera depends on distance, parts and budget.

Camera	How it works	Published accuracy
Intel RealSense D405	Passive stereo, close range	±2% at 50 cm; good on the wrist
Intel RealSense D435i	Active stereo, general purpose	RMS about 2 mm at 1 m
Orbbec Gemini 335L	Active stereo	0.8% at 2 m
Zivid 2+	Structured light	±0.2 mm at 1 m
Photoneo PhoXi L	Structured light	0.524 mm point to point