In brief: 3D robotic bin picking uses a depth-sensing vision system to locate randomly arranged parts, select an accessible grip and guide a robot through a collision-checked path. Success depends on part geometry, surface, overlap, container access, tooling and the required cycle.
Technical guidance by Oxford Industrial Automation. Updated 1 September 2026.
How a 3D bin picking robot works
The vision system captures the contents of the container and compares the scene with a trained or modelled part. Software estimates candidate poses, rejects inaccessible picks and sends a suitable target to the robot.
The robot approaches with a product-specific gripper, confirms the pick and transfers the part to a defined location. The cell must also manage re-scanning, failed grips, entangled parts, an almost-empty bin and a part that cannot be reached safely.
Part detection
3D data identifies candidate components despite varied position and orientation.
Grip planning
Geometry, surface and available contact determine vacuum, magnetic or mechanical tooling.
Robot path
Reach, wrist orientation, bin walls, other parts and the gripper shape constrain motion.
Downstream presentation
Picked parts are oriented into fixtures, conveyors, machines or inspection stations.
Bin picking trial inputs
| Part geometry | CAD, dimensions, mass, symmetry, holes, edges, deformability and entanglement risk |
|---|---|
| Surface | Colour, reflectivity, transparency, oil, dust, texture and permitted grip method |
| Container | Internal dimensions, wall shape, fill level, loading method and part density |
| Pick requirement | Required orientation, acceptable grip areas, destination and damage limits |
| Performance | Picks per minute, re-scan time, allowed failed picks and empty-bin threshold |
| Exceptions | Unrecognised parts, double picks, dropped parts, unreachable items and operator recovery |
Representative trials are essential for random bin picking
A feasibility test should use the real part surface, container, fill range and difficult arrangements. Clean CAD geometry alone does not show how glare, oil, overlap, casting variation or tangled components affect detection and grip.
Cycle testing should include scanning and re-scanning, not only the robot movement after a target has already been found. The final acceptance method should state both throughput and permitted intervention.