How to Make Robot Bin Picking Reliable: What Stable Part Grasping Requires ART-018

Stable picking starts with the full cycle, not a camera frame

In a random bin picking demonstration, the process often looks simple: a camera scans the bin, detected parts appear on screen, and the robot approaches one of them and removes it. In production, however, the useful outcome is not detection itself but a repeatable chain from a new bin state to a part that has been successfully picked and transferred onward.

Random bin picking is robotic picking from a bin where parts lie in arbitrary orientations, at different heights, and may partially occlude one another. In this kind of scene the system works with new geometry on every cycle: after one pick, neighboring parts may shift, expose a new surface, or make another grasp inaccessible.

A production cycle therefore requires several decisions at minimum: acquire usable 3D data, detect one or more parts, estimate their spatial pose, select a feasible grasp point, check collisions and reachability, execute the approach, confirm the grasp, exit the bin, and transfer the part to the next operation. If any link in this chain is unstable, “the camera sees the part” still does not mean the cell will operate autonomously.

The practical project question is: can the system consistently find at least one executable grasp in every expected bin state, from full to almost empty?

Visible, detected, and currently pickable are three different states

In a bulk-filled bin, part of each component's geometry may be hidden by neighboring parts. A 3D system works with the scene surface the sensor can actually acquire. For random bin picking, therefore, the quality of the part model is not the only issue; it also matters how much useful geometry remains visible across different orientations and partial occlusions.

A part can be detected correctly and still be unsuitable for the current pick. For example:

• the spatial pose estimate is correct, but the gripper would collide with a neighboring part;
• the accessible surface is too small for a vacuum cup or magnetic element;
• a finger gripper has no room to enter;
• the part is close to the bin wall and the tool body cannot pass;
• the intended grasp point is nominally free, but the robot cannot reach it in an allowable configuration;
• the part is trapped or hooked by neighboring parts so it cannot be extracted reliably.

That is why industrial logic should separate at least two questions: “what has been detected?” and “what can actually be picked now?” Pickit's official documentation separately describes grasp points, tool geometry, workspace constraints, and collision checking, so detection is only the beginning of the decision process.

For stability, it is preferable for the system to have several feasible candidates after one scan. Then a failed or inaccessible first option does not stop the entire cycle.

Test 3D scene quality on the worst surfaces and positions

A sensor does not operate in a vacuum. Gloss, dark surfaces, scale, oil, contamination, sharp edges, deep cavities, and mutual occlusion can all affect the quality of the acquired spatial geometry. With a particular camera technology, some conditions may be straightforward while others may require different sensor placement, different settings, or a different method of presenting the parts altogether.

That is why testing should use the real range of states rather than a clean “reference” part on a table:

• the top, middle, and bottom of the bin;
• parts near walls and in corners;
• top, bottom, and side orientations;
• partial occlusion by neighboring parts;
• the actual surface condition after the previous process step;
• the minimum and maximum expected bin fill levels.

One important point: seeing enough geometry in the top layer does not prove that the system is suitable for lower layers. Working distance, viewing angle, geometric shadowing, and tool access all change as the bin state changes.

Design the gripper together with the machine-vision system

In random bin picking, the gripper is not “the last mechanical component after the camera.” Its geometry directly determines how many detected parts become accessible for robotic picking.

Vacuum, magnetic, or finger grippers may each be appropriate for different parts. The choice depends on material, shape, mass, accessible surfaces, holes, roughness, allowable contact, and how parts actually lie in the bin.

For example, parts packed tightly against one another may leave no room for fingers while still presenting an open top surface for vacuum or magnetic picking. Conversely, a ring-shaped or spatially complex part may offer reliable internal or external clamping zones but very little area for a suction cup.

For a deep bin, the external geometry of the entire tool is especially important. The more of the gripper that has to enter between the walls and the parts, the more potential poses can be eliminated by collisions. Pickit's bin-picking documentation recommends minimizing the envelope of the part of the tool that enters the bin and modeling the tool shape for collision checking.

A long, thin tool is not an unconditional advantage either: its length affects stiffness, reachability, robot loading, and behavior during motion. The gripper should therefore be evaluated as a tradeoff between access into the bin, reliable part retention, and the kinematics of the entire cell.

One grasp point rarely covers every possible part orientation

If a part can lie on its top, bottom, or side, one nominal grasp point can make some detected orientations impractical.

For the part model, check:

• which surfaces can actually be used for picking;
• which tool orientations are allowable;
• whether symmetric equivalent poses exist;
• whether several grasp points can be defined;
• which grasp points should have priority;
• how much small pose or angular error the grasp can tolerate;
• whether the part needs to be regrasped after it is removed from the bin.

The last point matters especially when the next operation requires a strict orientation. Random bin picking may solve the task of “getting the part out,” but that does not mean the part immediately leaves the bin in a pose suitable for a machine, assembly, or inspection. In some cases the right architecture includes an intermediate reorientation station or secondary locating step.

Collision checking must cover both the grasp pose and the motion to it

There are at least three groups of potential obstacles inside the bin: the bin itself, other parts, and elements of the robot or gripper. After a successful pick there is another one: the picked part itself, which can strike a wall or neighboring object during exit.

The technical solution should therefore make the following explicit:

1. what is included in the collision model; 2. whether only the tool pose at the grasp point is checked or whether the approach and exit trajectory is also checked; 3. whether the geometry of the picked part is considered after the grasp; 4. how bin walls, nearby equipment, and other fixed obstacles are represented; 5. what happens when a calculated pose is unreachable for the specific robot.

This is not a formality. Official documentation for different systems explicitly distinguishes checking a grasp pose from full trajectory planning. The phrase “collision avoidance” in a quotation therefore needs to be unpacked into the actual level of functionality being provided.

Even a correct 3D pose does not guarantee that the robot can reach it: axis limits, flange orientation, proximity to a wall, cable position, and tool geometry can make some poses unreachable.

A failed pick should be a normal logic state, not an emergency

A real bin must account for situations where the selected part does not come out as expected: it shifts, turns out to be trapped by neighboring parts, the gripper does not confirm retention, or scene motion makes the previous detection result stale.

A stable cell needs a defined response to such events. For example:

• confirm a successful grasp with vacuum feedback, finger-position sensing, magnetic monitoring, or another method suitable for the specific tool;
• if the grasp is not confirmed, retreat safely;
• avoid repeating the same failed grasp indefinitely when other candidates exist;
• rescan the scene when parts have moved noticeably;
• define a rule for the state “the bin is not empty, but no feasible candidates are available”;
• distinguish the need for rescanning from a genuine need for operator intervention.

Pickit, for example, provides functions for re-detection and temporarily avoiding parts that have just failed to be picked. But the specific failed-pick recovery logic depends on the robot integration and the production process. It needs to be tested as seriously as detection itself.

Test the bottom, corners, and walls of the bin

A full bin often provides many candidates near the top. More informative states appear toward the end of the batch, when parts remain near the walls, in corners, or on the bottom and the available approach space becomes smaller.

This is where problems that are easy to miss in a demonstration become visible:

• the robot flange or gripper body no longer fits into the required zone;
• a convenient grasp point faces the wall;
• the part is visible, but the trajectory is unreachable;
• several parts are hooked together or tightly jammed;
• the scene changes after a pick attempt and requires a new scan;
• the system has no clear rule for when the bin is considered finished and when operator intervention should be requested.

An acceptance test should therefore include not only a quick start with a full bin but also emptying the bin down to the real end state.

Random bin picking cycle time changes with the scene

A single “pick time” value says little without context. In some cycles the camera immediately finds several convenient candidates, planning is short, and the robot performs a simple vertical approach. In another bin state, the system may need to rescan, use a more complex orientation, make a longer motion, or reject several inaccessible poses.

Production evaluation should therefore measure the full cycle:

scan → detection → select a feasible candidate → planning → move to the part → grasp → confirm → exit the bin → transfer/reorientation → ready for the next cycle.

Track the frequency of rescans, failed picks, and manual interventions separately. These events often explain the difference between an attractive short demonstration cycle and the real throughput of the cell.

There is no universal cycle time for random bin picking: it depends on the robot, sensor, part, bin, gripper, algorithms, required output orientation, and the current scene state.

When random bin picking has good prerequisites

This architecture is particularly logical when parts genuinely arrive unordered and additional mechanics for precise orientation would be complex or inflexible.

Good prerequisites include:

• the part geometry provides enough recognizable surface in different orientations;
• the sensor produces usable 3D data on the real material surface;
• most orientations offer at least one practical grasp point;
• the tool can enter the bin without an excessive number of collision exclusions;
• the parts do not systematically jam or mechanically hook together;
• the robot can reach the required working area of the bin;
• a failed pick does not stop the cell without a clear recovery path;
• the next operation can accept the output pose, or a reorientation step is provided;
• the natural variability of cycle time is compatible with the required production takt.

This is not a checklist that guarantees success. It is a set of prerequisites that justifies moving on to a real test.

When it is better to structure the part presentation first

Random bin picking does not have to be the best answer simply because parts currently arrive in bulk. Sometimes it is cheaper and more stable to reduce the disorder before the robot starts searching for a part.

Compare an alternative architecture when:

• parts routinely intertwine, wedge into one another, or lift in groups;
• the accessible grasping area is too small for most random orientations;
• the next operation requires a very specific pose, so almost every part must be reoriented anyway;
• a narrow and predictable production takt is required, while rescanning or complex trajectories create unacceptable variability;
• the product range is small and stable, so mechanical locating, a cassette, magazine, step feeder, or semi-structured presentation may be simpler;
• manual or automatic pre-separation dramatically increases the number of feasible grasps.

The goal is not to “defeat” disorder with a more complex algorithm, but to choose the architecture with the least total uncertainty across the full production cycle.

How to test the system before deciding to buy

A proper test should prove the viability of the whole scenario, not one successful pick.

Use real parts in their actual surface condition and a bin as close as possible to the production version. If the bin has not yet been selected, test several realistic geometries rather than an abstract open platform.

Then run through a series of different states. Deliberately create difficult cases rather than selecting a “convenient” pile.

Store the test result as a log rather than one average number: bin state, selected part, why other candidates were rejected, grasp success/failure, need for rescanning, full-cycle time, and the reason for manual intervention if one occurred.

What data to give the integrator at the start

The earlier the integrator sees the real part and real presentation method, the lower the risk that the camera, robot, and gripper will be selected independently of one another.

For an initial assessment, prepare:

• a 3D/CAD model of the part, if available;
• actual part samples;
• material, mass, overall dimensions, and surface condition;
• photos or video of the real bulk arrangement;
• internal dimensions, depth, and shape of the bin;
• minimum and maximum expected fill levels;
• required part orientation after picking;
• the next operation and available transfer zone;
• desired production takt, including how much variability is acceptable;
• the states where manual intervention is acceptable and the states where autonomous recovery is required.

This is enough to start the discussion with the cell architecture and acceptance test rather than with a standalone camera model.

Next step: test your scene, not just the camera