Industries / Robotics & Industrial AI
Robotics & Industrial AIPerception data for machines that pick, place and move.
A robot acts on what its perception model says is there. We prepare and validate the labels behind that: parts and objects, arm and object keypoints, and the sequences that show how a task unfolds.
The data problem.
The difficulty is geometry, and what the camera cannot see.
Touching, stacked parts
Bins of identical items, where instance boundaries are the hard part.
The gripper hides the object
Hidden keypoints still need a call.
Shiny, dark and clear surfaces
Metal and clear packaging confuse depth sensors.
Every cell is different
Lighting and mounting change between stations.
What we label.
- RGB images and video, fixed and arm-mounted
- LiDAR and 3D scans of robots and cells
- Task sequences
- Part detection with boxes and instance masks
- Keypoints on arms, tools and objects, with visibility flags
- Tracking parts and the arm
- State tags: picked, placed, failed
How it is checked.
Approaches we use or evaluate: segmentation and keypoint models propose labels; depth plane fitting separates objects from bins. Reviewers decide, by written rules:
- Hidden keypoints: inferred or marked occluded
- Boundaries between touching parts
- The class of near-identical part variants
- Consistent IDs and states through the sequence
A palletising cell, labelled.
A palletising robot: five arm keypoints, the gripper, crates and pallet stack marked, one occluded crate routed for review.
Point AI’s role.
The data operation, not just the labelling queue.
- Calibration: a pilot on your cells sets parts, keypoints and tolerances
- Operations and review by our employed team
- Delivery in Point AI Studio: review in the browser, export or API
Have a challenging data problem?
Tell us what you’re building and what data you have. We’ll suggest a practical way to process, check and deliver it.