Bin picking is the robotic task of locating, grasping and extracting objects that lie in random positions and orientations inside a container, using 3D vision, grasp planning and collision-aware motion. In robotics, it is one of the most demanding forms of manipulation because the robot must cope with clutter, occlusion and contact that structured feeding systems are designed to avoid.
Bin picking removes the need for part-specific feeders, trays and manual loading, which makes it attractive for machine tending, logistics and assembly. Its difficulty explains why it was long called a holy grail of industrial automation and why reliable deployments still require careful engineering.
How a bin picking system works
A bin picking system runs a perceive, plan, act loop for every part it extracts. The cycle starts with a 3D scan of the bin, continues with object detection and grasp selection, and ends with a collision-free motion that lifts the part and places it at a defined target.
| Stage | Purpose | Typical techniques |
|---|---|---|
| Scene capture | Acquire geometry of the bin contents | Structured light, stereo, time-of-flight |
| Object localization | Find candidate parts and their poses | CAD matching, point cloud segmentation, deep learning |
| Grasp planning | Choose a reachable, stable grasp | Grasp databases, learned grasp quality scores |
| Motion planning | Approach and extract without collisions | Sampling planners, collision checking against bin walls |
| Verification and placement | Confirm pick and orient the part | Vacuum sensing, secondary vision, regrasp station |
The robot typically scans once per pick or once per several picks, because each extraction disturbs the pile. Systems that reuse a scan for multiple picks gain speed but risk acting on outdated scene information when parts shift.
Levels of structure in bin picking
Practitioners distinguish three levels of bin picking according to how much order the parts have. The level determines the sensor, the algorithm and the achievable cycle time far more than the choice of robot arm.
| Level | Part arrangement | Vision needed | Typical cycle time |
|---|---|---|---|
| Structured picking | Parts in layers, trays or dividers | 2D or simple 3D | 2-5 s |
| Semi-structured picking | Loose but similarly oriented parts | 2.5D or 3D | 4-8 s |
| Random bin picking | Fully chaotic pile, overlapping parts | Full 3D with pose estimation | 5-15 s |
A related distinction separates known-object picking, where CAD models of the parts exist, from unknown-object picking in logistics, where the system must grasp items it has never seen. The second case relies on model-free grasp detection that scores surface patches for suction or finger contact rather than matching a model.
Grippers, sensors and their trade-offs
The end effector decides which parts can be picked at all. Vacuum cups need a flat, sealed surface; finger grippers need free space around the part; magnetic grippers work only on ferrous material and tend to lift several parts together.
| Gripper | Strengths | Weaknesses |
|---|---|---|
| Vacuum suction | Compact, tolerant of pose error | Fails on porous or curved surfaces |
| Parallel fingers | Precise, defined part orientation | Needs clearance, collides with neighbors |
| Magnetic gripper | Fast, simple for steel parts | Double picks, residual magnetism |
| Soft or adaptive gripper | Handles varied shapes gently | Lower placement accuracy |
Sensor choice involves a similar balance. Structured-light cameras offer sub-millimeter accuracy at short range but struggle with shiny metal and black plastic, which produce reflections or missing data. Mounting the camera on the robot wrist allows multiple viewpoints, while a fixed overhead sensor can scan the next pick in parallel with robot motion, shortening cycle time.
How bin picking relates to other manipulation tasks
Bin picking is a specialized form of pick and place in which the source position is unknown and cluttered. It builds directly on several neighboring capabilities rather than replacing them.
- Pick and place assumes known or easily detected positions; bin picking adds perception under clutter.
- Grasp planning supplies candidate grasps; bin picking filters them by reachability and collision risk.
- Inverse kinematics checks whether each grasp pose is reachable within joint limits.
- Motion planning with collision checking keeps the arm and gripper clear of bin walls and other parts.
- Dexterous manipulation becomes relevant when parts must be reoriented in hand after extraction.
Because the grasp seldom leaves the part in its final orientation, many cells add a regrasp or orientation station, where the robot sets the part down on a flat fixture and picks it again precisely. This step trades a few seconds for the placement accuracy required by machine tools and assembly presses.
Failure modes and practical limits
Real deployments rarely fail on the first pick; they fail on the last ten percent of the bin, where parts lie flat against walls and corners. Planning for bin emptying performance is therefore as important as average success rate.
| Problem | Cause | Countermeasure |
|---|---|---|
| Parts left in corners | Gripper collides with bin walls | Angled approaches, tilting bins, slim tooling |
| Entangled parts | Springs, rings, hooks interlock | Shake routines, dedicated separation step |
| Missing 3D data | Reflective or dark surfaces | Polarized or multi-exposure capture |
| Double picks | Magnetism, oil films, suction leaks | Weight or vision check before placement |
| Long cycle times | Planning after every scan | Overlap scanning with robot motion |
A common mistake is to evaluate a system only on its pick success rate, often quoted at 95 to 99 percent. The figures that decide economic viability are mean cycle time including retries, the share of parts that remain in the bin, and the frequency of operator interventions per shift.
The short version
Bin picking means teaching a robot to reach into a messy box of parts and take them out one at a time. A person does this without thinking. A robot must first see the jumble in three dimensions. Then it decides which part is easiest to grab. Finally, it moves its arm without hitting the box or the other parts.
The payoff is simple. Factories no longer need someone to sort parts by hand. They also avoid expensive machines that line parts up for the robot.
Where it makes a difference
A supplier of automotive brackets runs two CNC machining centers that previously required an operator to load steel blanks from a wire mesh bin by hand. The company installs a random bin picking cell with a six-axis robot, a fixed structured-light sensor above the bin and a magnetic gripper with a small vacuum cup as backup.
The vision system matches the blank CAD model to the point cloud, the planner rejects grasps that would collide with the bin walls, and the robot places each blank on a regrasp fixture before loading the machine chuck. The overhead sensor scans the next pick while the robot is loading, which keeps the average cycle near 9 seconds.
After tuning the approach angles for corner parts, the cell empties about 98 percent of each bin without assistance. One operator now supervises both machines and refills bins every two hours instead of loading parts continuously.