Common Requirements and Troubleshooting
About 1818 wordsAbout 6 min
7. Common Requirements and Troubleshooting
First, use the fixed samples in Operation and Acceptance to establish reproducible results. Then follow this chapter to address order, cycle time, or abnormal behavior. After each change, retest both the problem sample and normal samples.
7.1 Adjust the Picking Order for Neatly Arranged Targets
7.1.1 Process by Row, Column, or a Specified Order
- First determine how the cylinder arrangement direction corresponds to the sorting coordinate axes.
- For regular rows and columns, use the general picking strategy to group and sort instances. Check whether the grouping step size merges adjacent rows into one group or splits a single row.
- If targets arrive in fixed positions and those positions must be processed in a specified order that the general strategy cannot reliably achieve, consider a custom picking strategy.
- Validate with samples that have missing targets, edge offsets, and different arrangements; do not check only a fully loaded arrangement.
- Check whether sorting is also configured at the Pick Point stage, which could make the final output order differ from the instance order.
For coordinate systems and combination restrictions of custom strategies, see the Instance Sorting Configuration Guide. For sorting output points, see the Pick Point Sorting Configuration Guide.
7.1.2 Prevent Failed Positions from Being Skipped
When the process requires strict sequential handling, consider Pose Calculation Exception Alarm. When enabled, poses are calculated in instance sort order, and an alarm is raised if 3D pose estimation fails. When disabled, the current instance may be skipped and calculation may continue with subsequent instances.
This option addresses pose calculation failures. It does not replace a check for positions that the model fails to identify entirely. If the process requires every position to be present and handled, verify how the actual Task logic responds to missing targets and missed detections.
For configuration details, see the Vision Calculation Configuration Guide.
7.2 Use One Capture for Multiple Picks
If cylinder positions remain unchanged during consecutive picks and a preceding pick does not disturb the remaining Targets, consider using one capture for multiple picks:
- First complete identification and robot integration acceptance for one capture per pick.
- In the vision calculation configuration, set the required Number of Poses to Calculate and Number of Instances to Send, then enable one capture for multiple picks.
- Check how subsequent Robot requests interact with cached vision results.
- During consecutive picks, confirm that the remaining cylinders have not shifted, tilted, or been moved by the gripper.
- After replenishment, tray replacement, manual intervention, or Target movement, clear the old results according to the communication protocol and capture again.
Conditions for Using Cached Results
The cached Pick Point coordinates correspond to the Scene at the time of capture. If removing one cylinder causes adjacent cylinders to move, do not continue using the old coordinates without validation. One capture for multiple picks also does not mean sending all instances each time.
For configuration and cache interactions, see the Vision Calculation Configuration Guide and Robot Configuration Guide, respectively.
7.3 Optimize Vision Cycle Time
Use actual log timings to locate bottlenecks. Do not assume that the same stage dominates every Scene.
| Time-consuming stage | Check first | Also verify |
|---|---|---|
| Image and instance identification | Whether the ROI is too large and whether auto enhancement includes ineffective combinations | Targets at the edges and within permitted incoming variations are not missed |
| Point Cloud and pose calculation | Whether too many invalid instances are processed and whether the template point count and Fine matching mode are appropriate | Effective features, matching accuracy, and orientation do not deteriorate |
| Pick Point processing | Number of candidate Pick Points and complexity of the collision model | Available picking methods and collision checks are not lost |
| Data saving | Whether all debugging data is always saved | Information needed to investigate abnormal behavior remains available |
After the basic process passes acceptance, consider Save Original Images Only (Ultra-Fast Mode) or saving data as needed to reduce write overhead. Do not disable necessary angle, region, or collision filtering solely to improve cycle time.
For template downsampling, see the Point Cloud Template Creation Guide. For data saving and calculation counts, see the Vision Calculation Configuration Guide.
7.4 Troubleshoot by Processing Stage
7.4.1 No Instances Detected or Instances Missed
- Confirm that the current Target, Visual Model, and Pattern mask are correct.
- Examine the original image. Check imaging, reflections, and whether the Target is within the 2D/3D ROI.
- Compare images before and after preprocessing to confirm that cylinder edges or orientation features have not been cropped out.
- On fixed samples, adjust the scale and minimum Confidence Threshold incrementally while tracking false detections.
- If a single Parameter cannot cover normal variations, then consider auto enhancement.
- If results remain unstable, collect approved representative samples and consider updating the model. Do not rely on an excessively low Threshold over the long term.
See the 2D Instance Segmentation Configuration Guide and 2D Preprocessing Function Configuration Guide.
7.4.2 Instances Detected but No Poses Produced
- Check whether instances have been merged, truncated, or incorrectly include the background.
- Check whether the generated Point Cloud is complete and whether 3D preprocessing filters out too much data.
- Verify that the mesh, keypoints, Visual Model, and template correspond to one another.
- Examine the logs to distinguish keypoint Coarse Matching failures from Point Cloud Fine matching failures.
- For Coarse Matching failures, check the keypoints and model first. For Fine matching failures, check the visible template surfaces, alignment, and valid point pairs first, then adjust the relevant Parameters.
See Point Cloud Template and Target Configuration and the General Pose Estimation Configuration Guide.
7.4.3 A Pose Is Produced but Is Offset, Flipped, or Unstable
- Consistent offset: First check dimension units, the template coordinate reference, and the Pick Point position. Then distinguish vision-side issues from Robot-side issues.
- Front-back or end-to-end flip: Check the keypoints, corresponding template surfaces, and effective orientation features. Do not determine correctness from the matching score alone.
- Variation across repeated captures: Compare original Point Cloud quality, changes in reflections and occlusion, and whether Fine matching is drawn toward adjacent Targets or the background.
- Only the Robot position is abnormal: Check calibration, TCP, coordinate systems, units, and rotation representation. Do not first compensate by offsetting the vision Pick Point.
If a rotationally symmetric Target has genuine ambiguity around its axis, consider axis rotation pose adjustment as described in the General Pose Estimation Configuration Guide, and observe its restrictions on combining functions. It does not need to be enabled for every cylinder.
7.4.4 The Pose Is Correct but No Pick Point Is Output
- Confirm that a Pick Point has been added manually to the Target and saved to the currently selected Target.
- Examine the result counts before and after each filter to locate the first function that excludes the Pick Point.
- Check the matching score, angle axis and positive or negative direction, and ROI workspace or picking region boundaries separately.
- If collision filtering is enabled, check the gripper model coordinates and dimensions. Do not substitute disabling collision detection for analyzing the cause.
- Check whether the pose calculation count, send count, or picking logic limits this output.
See the Pick Point Filtering Function Configuration Guide.
7.4.5 A Lower-Layer or Occluded Cylinder Is Output First
- First confirm whether upper-layer Targets were missed; do not change sorting alone.
- Verify the coordinate system, axis direction, and grouping step size used for height sorting.
- Check whether upper-layer Targets were skipped after pose calculation failed.
- Configure occlusion filtering and exception alarms according to the process, and validate with samples in which upper-layer Targets have not been removed.
- Resume Robot integration only after the system retains only pickable Targets and its response to failures meets requirements.
Do not apply specialized filtering methods for interlayer paper indentations or center holes from other Target cases directly to cylinders. Make decisions based on the actual geometry and supported functions.
7.5 Prepare Reproduction Data and Check Materials
If further analysis is needed, collect the software version, problem description, general sample number, configuration version, corresponding historical data, and runtime logs. State the expected and actual results.
Before sharing, check the document body, filenames, archive directories, and screenshots. Remove or replace sensitive information such as customer names, project codes, people and contact details, internal addresses, account credentials, device identifiers, and QR codes. For attachments containing real Target models or on-site images, also confirm the permitted sharing scope and purpose; they are not default downloads for this public case.
Return to Case Overview · Previous Chapter: Operation and Acceptance