Grasp validation and failure recovery

1. Engineering problem

Picking fails in several distinct ways: nothing was grasped, the wrong thing was grasped, the grip is weak, or the object slips while being moved. Each needs a different response, and each can be detected from the gripper’s own measurements.

2. When is this needed?

• Building a reliable pick-and-place cycle.
• When failures are silent.
• Adding retries.

3. Selection and design decisions

Signals: finger width after closing (a fully closed gripper means nothing was grasped; a width that differs from the expected object width suggests the wrong object or pose), gripper force or motor current (below a minimum means a weak grasp), and width again after lifting (an opening width means slip). Vacuum grippers use the vacuum level in the same way.

4. Advantages and disadvantages

• More checks catch more failures and add time to each cycle.
• Retry limits avoid infinite loops and need an escalation (ask for help).
• Sensors in the gripper cost more and pay back in reliability.

5. Requirements and constraints

• Expected width and minimum force for each object type.
• A retry limit and an escalation path.
• All failures are logged with the object and pose.

6. Architecture

close ─▶ check width and force ─▶ lift ─▶ check width again ─▶ move ─▶ place
          failure at any check ─▶ classified ─▶ retry or escalate

7. Implementation

1. Define expected width and minimum force per object.
2. Add checks after closing and after lifting.
3. Classify failures (below) and map each to a recovery.
4. Limit retries and add an escalation.
5. Log every failure and review the statistics.

def classify(width_closed, width_expected, tol, force, force_min, width_after_lift):
    if width_closed <= 0.002:
        return "missed: fingers closed on nothing -> re-detect and retry"
    if abs(width_closed - width_expected) > tol:
        return f"wrong object or pose: width {width_closed * 1000:.0f} mm, expected {width_expected * 1000:.0f} mm -> release and re-grasp"
    if force < force_min:
        return "weak grasp: force below the minimum -> regrasp with more force or another pose"
    if width_after_lift is not None and width_after_lift - width_closed > 0.003:
        return "slipped during lift: width opened -> lower speed or regrasp"
    return "grasp ok"
cases = [(0.001, 0.05, 0.005, 0.0, 10.0, None), (0.032, 0.05, 0.005, 12.0, 10.0, None), (0.050, 0.05, 0.005, 4.0, 10.0, None), (0.050, 0.05, 0.005, 14.0, 10.0, 0.056), (0.051, 0.05, 0.005, 14.0, 10.0, 0.051)]
for c in cases:
    print(classify(*c))

Printed (checked by the site’s tests):

missed: fingers closed on nothing -> re-detect and retry
wrong object or pose: width 32 mm, expected 50 mm -> release and re-grasp
weak grasp: force below the minimum -> regrasp with more force or another pose
slipped during lift: width opened -> lower speed or regrasp
grasp ok

Reading it: width 1 mm means a missed grasp; 32 mm where 50 mm is expected means the wrong object or pose; correct width with 4 N below a 10 N minimum is a weak grasp; the width opening by 6 mm after lifting is a slip; and 51 mm with 14 N and no change is a good grasp. The thresholds and values are illustrative.

8. Real examples

• Kitting cell: two retries with a re-detection, then the cell calls for help.

9. Failure scenarios

• No check after lifting, so slip is invisible.
• Unlimited retries on a mis-calibrated camera.
• Not logging the failure type.

10. Troubleshooting

1. Group failures by type and object.
2. Check the width and force thresholds on good grasps first.
3. Check calibration if misses are systematic.

11. Validation

• The classifier prints the lines shown (executed by the site tests).
• Failure statistics from a long run show a known rate per type.

12. Engineering checklist

• Expected width and force
• Checks after close and lift
• Failure classes mapped
• Retry limit and escalation
• Logging

13. Further references

• MoveIt pick and place tutorial: moveit.picknik.ai, read 2026-11-06 (moveit_msgs/Grasp: grasp_pose, pre_grasp_approach, post_grasp_retreat).

Related on FixTheRobot