Crankshaft Bin Picking: Acceptance Tests That Matter
Crankshaft bin picking should be approved against a declared range of pile states and recovery cases, not a single successful pick. The critical result is a stable grip that clears the bin, preserves orientation, reaches the downstream datum, and exits safely when no acceptable pose exists.
This guide uses a Pick-Recover Envelope. EVST applies it to separate measurable production states from a vision demonstration. It covers scan, pose choice, grip, rim clearance, transfer, placement, and depletion behavior; it does not claim that every random pile is pickable or publish a universal cycle for an untested bin and part family.
Crankshaft bin picking starts by defining the pile population
The word “random” is not an acceptance specification. Describe the permitted crankshaft variants, mass and center of gravity, surface condition, oily or dry state, bin dimensions, fill height, liner, allowable overlap, and any orientation deliberately excluded upstream. Include the depleted-bin condition; many systems see their most difficult occlusion and reach geometry near the bottom.
Build a representative pile set rather than arranging only visually convenient parts. The set should contain dense layers, crossed journals, end-to-end contact, parts close to the wall, unstable top pieces, and the least favorable state the infeed method is expected to create. Label the samples or preserve the loading method so a failed trial can be repeated.
The acceptance envelope needs an explicit outside. A bent container, mixed foreign object, unknown part, unreachable corner, or pile outside the modeled height may require a hold rather than a pick attempt. Defining that boundary is safer and more useful than forcing the planner to return some pose for every scan.
Make perception output auditable
A 3D sensor produces data; the application must convert it into a pick decision. Record the sensor pose, calibrated relationship to the robot and bin, field of view, occlusion zones, exposure or surface limitations, point-cloud checks, part model, and the criteria used to rank poses. A confidence value is meaningful only when its response to known bad scenes has been tested.
The perception record should distinguish “part detected,” “candidate pose generated,” “candidate collision-free,” and “candidate accepted for execution.” These are different states. If the system filters a pose because the gripper cannot approach, the bin rim blocks extraction, or the downstream orientation cannot be reached, that rejection is useful evidence—not a vision failure to hide.
Challenge calibration and scene validity intentionally. Move the bin within its allowed locating tolerance, vary the fill level, introduce reflective or oily surfaces representative of production, cover features, and interrupt the scan. Confirm that invalid or stale data cannot release robot motion. After a bin exchange, require a fresh identity and scene state before planning.
| Scene condition | Expected planner result | Required cell action |
|---|---|---|
| Valid pose with stable grasp and clear extraction | Executable candidate | Pick, verify grip, and monitor rim clearance |
| Detected part but blocked approach | Rejected candidate | Evaluate another pose without moving blindly |
| No collision-free stable pose | Defined no-pick state | Rescan, request controlled bin action, or call for intervention |
| Invalid scan or calibration state | Perception fault | Hold robot and preserve the bin state for diagnosis |
Design the gripper around mass, journals, and release
Crankshaft geometry makes grip choice more than a finger-shape exercise. Map allowable contact surfaces, journal protection, mass distribution, center of gravity, potential rotation inside the fingers, and how adjacent parts can be trapped or lifted. Include the gripper, adapters, sensors, hoses, and cables in payload, inertia, wrist-moment, and reach checks.
Grip proof should correspond to the failures that matter. Position switches may confirm finger travel but not necessarily one secured crankshaft. Pressure can reveal gross loss but may miss a part caught in an unintended contact. A secondary observation or motion-limited lift test may be appropriate, provided the method is validated and does not make an unstable load more hazardous.
Release deserves the same detail. Define the downstream locator, allowed approach, seating feature, orientation check, and response when the fixture is occupied or the part does not settle. A successful bin extraction is not a completed cycle if the crankshaft cannot be placed to the datum required by the next operation.
Treat bin-rim clearance as a dynamic test
The most demanding collision can occur after the fingers close. A long crankshaft may rotate under gravity, shift within the grip, or swing as the wrist changes orientation. Meanwhile the surrounding pile can settle into the volume the planner believed was clear. Model the part, gripper, cable package, bin wall, and a conservative representation of neighboring parts through the full lift.
Use a deliberate extraction strategy: slow initial separation, observation or grip confirmation at a safe height, defined wrist orientation near the wall, and a collision response that does not drop the part back into an unstable pile. The correct sequence depends on geometry; the principle is to validate clearance under load rather than from the planned tool point alone.
Challenge candidates near each wall and corner, including the lowest permitted layer. Record minimum observed clearance or another agreed evidence method, not just whether the program avoided a protective stop. Protective stops are safeguards or protective functions; they are not a substitute for an acceptable planned trajectory.
Prove the no-pick state before claiming autonomy
Every random-bin system eventually sees a scene with no approved move. The important question is whether the cell recognizes it early and enters a bounded recovery. Options may include a rescan after settling, a validated bin agitation step, a change of sensor view, a controlled request for operator intervention, or removing the bin from service. The project must choose and test the allowed options.
Set retry limits and record why each attempt failed. Endless rescanning can waste cycle time without changing geometry; repeated contact can make the pile less stable. If an operator enters or manipulates parts, define the safe access mode, hazardous-energy controls, reset position, visibility, and the new scan required after the scene changes.
No-pick rate is not a universal product number. It depends on pile creation, fill level, part variants, surface behavior, sensor layout, gripper geometry, collision model, and planner rules. If the project reports it, state the bin population, sample size, method, and confidence boundary used for the trial.
Keep downstream availability in the pick decision
The interface state includes bin present, scan valid, grip confirmation available, transfer zone clear, placement fixture ready, and reject or hold location accessible. The planner should not lift a heavy part if the only safe destination is blocked. Reserve the downstream handoff before committing to a pick, or provide a validated intermediate location.
Part identity and orientation must survive the transfer. If the system handles variants, bind the selected model and pose to the physical part until placement verification completes. After a communication interruption, do not assume that the active record still describes the crankshaft in the gripper.
At placement, confirm the locating state the downstream process actually needs. A robot-reached signal is insufficient when the part can bridge a locator, rotate after release, or remain in the fingers. Use the specified physical evidence and route uncertain placements to a recoverable state before allowing the next machine action.
Include heavy-part and recovery hazards
Relevant hazards include suspended mass, dropped parts, bin-edge collision, unstable piles, pinch points, and manual access for recovery. The risk assessment covers normal operation plus bin exchange, sensor cleaning, teaching, gripper service, jam clearing, dropped-part retrieval, and work below or near a suspended load.
ISO 10218-2:2025 addresses the industrial robot application and cell, while OSHA’s robot guidance includes end effectors, controls, power sources, sensors, interfaces, and application hazards. The installation still needs local legal review and an application-specific risk-reduction design. Robot stop functions do not replace mechanical retention or a controlled recovery method for a heavy crankshaft.
Test power loss, air loss, grip-signal disagreement, emergency stop, and restart with a part in each credible location. State whether the gripper retains the part, where it can be lowered, and which manual actions are prohibited until energy and load are controlled.
Measure depletion, not only the first layer
Segment scanning, candidate generation, approach, grip proof, initial lift, rim clearance, transfer, orientation check, placement, and rescan. Observe the full range from high fill to the declared low-fill limit. Report waiting for the downstream fixture and recovery separately from active robot motion.
A single “average pick time” can hide the state that constrains output. Early picks may be easy while the last layer requires more planning or intervention. Use distributions by fill band or another transparent method, include no-pick and recovery events, and avoid deriving production throughput from an edited video.
Build a Pick-Recover acceptance campaign
The trial matrix should include representative piles, wall and corner candidates, low-fill scenes, calibration shifts inside the allowed tolerance, invalid scans, uncertain grips, rim contact risk, blocked placement, communication loss, and controlled recovery. Retain scene data, selected pose, collision decision, grip evidence, part identity, placement result, and disposition for failed tests.
Provide these inputs to EVST:
- representative crankshafts and allowed variants
- mass, center of gravity, and protected surfaces
- bin drawings, fill states, and exchange method
- downstream locator and required orientation
- recovery permissions and operator-access method
- target cycle envelope and evidence format
The review can use NIST’s measurement-oriented approach to turn “pick reliably” into observable requirements. Any claim about pick rate or recovery remains blocked until the sample population and method are agreed.
Frequently asked questions
Is a high vision confidence enough to execute a crankshaft pick?
No. The candidate also needs a feasible gripper approach, stable contact, payload and wrist capacity, collision-free extraction, and a reserved destination. Confidence describes one part of perception; it does not prove the mechanical or downstream states.
What should happen when no safe pose exists?
The system should enter a declared no-pick state. It may rescan, perform an approved scene-changing action, request a bin exchange, or require controlled intervention. The option, retry limit, safeguarding, and fresh-scene requirement must be tested before production.
Can one gripper cover multiple crankshaft models?
Possibly, but coverage has to be proven across geometry, mass, center of gravity, allowable contact, neighboring-part entanglement, extraction clearance, and downstream orientation. A common nominal diameter alone is not enough to approve the family.
Does one emptied bin prove the cell is production-ready?
No. One bin does not represent the declared pile population, contamination, variants, calibration tolerance, sensor limitations, blocked downstream states, or repeatable recovery. Acceptance should use a documented campaign and retain failed scenes as evidence.
Conclusion
Crankshaft bin picking becomes an industrial process when pick success and no-pick recovery are both designed outcomes. Share the parts, bins, pile range, gripper limits, handoff datum, recovery rules, and measurement method with EVST. The resulting Pick-Recover Envelope should tell operators exactly what the cell can accept, reject, and recover—without pretending every random pile is solvable.
Related EVST reading
- Existing 3D-vision bin-picking cell reference
- Industrial robot architecture and application range
- Payload, reach, and robot-type selection
References
- ISO 10218-2:2025 — Industrial robot applications and robot cells — used for integration, commissioning, and application-level safety boundaries.
- OSHA Technical Manual: Industrial Robot Systems and System Safety — used for system hazards, controls, end effectors, and safety evaluation.
- NIST Robotic Systems for Smart Manufacturing Program — used for measurable performance requirements and verification methods.