Semiautomatic dimensional screening of plastic scintillator cubes using image analysis and robotics

The authors developed and validated a semiautomatic system combining robotics, high-resolution imaging, and image analysis to efficiently screen 1 cm³ plastic scintillator cubes for future neutrino detectors, achieving 10 μm precision, over 80% consistency with manual screening, and a 3.1% rejection rate.

The Gist

Imagine you are building a massive, 3D puzzle to catch invisible particles called neutrinos. This puzzle, called the SuperFGD, is made of nearly 2 million tiny plastic cubes, each the size of a sugar cube.

To make this puzzle work, you need to thread thin, flexible fibers (like fishing line) through holes drilled in the center of every single cube. If the holes aren't perfectly aligned, the fibers will get stuck, bend, or snap, ruining the detector.

The Problem:
In the past, humans had to check these millions of cubes by hand. They would stack 15x15 cubes and try to push metal rods through the holes. If a rod got stuck, the cube was tossed out.

• The issue: It was slow (taking 16 seconds per cube), tiring, and subjective (one person might think a rod is "stuck," while another thinks it's fine).
• The goal: Build a robot that can check these cubes faster, more accurately, and with a clear "yes or no" answer.

The Solution: A Robot Camera Team

The authors built a machine that acts like a super-precise quality control inspector. Here is how it works, broken down into simple steps:

1. The "Turntable" (The Mechanical System)

Imagine a spinning pizza turntable with eight little platforms on it.

• A human places a plastic cube on a platform.
• The turntable spins. As it spins, the cube rolls slightly into a corner, ensuring it sits in the exact same spot every time (like a ball settling into a cup).
• The turntable stops at two specific spots. At each stop, six high-powered cameras snap photos of the cube from all angles, just like a 360-degree security camera system.

2. The "Digital Eye" (Image Analysis)

Once the photos are taken, a computer program (the brain) analyzes them. It doesn't just look; it measures with microscopic precision.

• The Ruler: It measures the cube's size to the nearest 10 micrometers (that's 1/100th the width of a human hair).
• The Bump Detector: It looks for tiny bumps or rough edges that might prevent the cube from stacking perfectly.
• The Hole Hunter: It finds the three holes drilled in the cube and measures exactly where they are.

The Analogy: Think of this like a bouncer at a club. The bouncer checks your ID (size), looks for a fake tattoo (bumps), and makes sure you are standing in the right spot (hole position). If you don't fit the criteria, you get kicked out.

3. The "Sorting Robot" (The Robotic Arm)

In the first version of the machine, cubes that failed the test were just dropped into a "trash" bin. But the scientists realized something important: Not all "bad" cubes are equally bad.

Some cubes have holes that are shifted slightly to the left. Others are shifted slightly to the right. If you mix them all up, the fibers won't fit. But if you group the "left-shifted" cubes together and the "right-shifted" cubes together, the fibers will slide right through!

• The Upgrade: They added a 6-axis robotic arm (like a human arm with a wrist and fingers).
• The Job: Instead of just dropping cubes, the robot gently picks up each cube, checks its "personality" (where its holes are shifted), and places it into one of 48 different bins.
• The Result: It sorts the cubes so that every row in the final detector has cubes with matching hole shifts.

The Results: Why This Matters

• Speed: The robot checks a cube in about 15 seconds. Humans took over 16 seconds just for a simple check, and they couldn't do it for millions of cubes without getting tired.
• Accuracy: The robot is incredibly consistent. It agreed with human experts 80% of the time, but it never got tired or bored.
• Efficiency: By using the robotic arm to sort the cubes into 48 groups, they saved thousands of cubes that would have been thrown away. The "rejection rate" (cubes thrown in the trash) dropped from a high number down to just 3.1%.

The Big Picture

This paper isn't just about plastic cubes; it's about how we build the future of science. As particle physics experiments get bigger and more complex (like the DUNE experiment in the US), we can't rely on humans to check every single part. We need smart, automated systems that can see tiny details, make logical decisions, and sort millions of parts with the precision of a surgeon.

This system is the blueprint for building the next generation of giant, universe-mapping machines.

Technical Summary

1. Problem Statement

Large-scale particle physics detectors, such as the SuperFGD in the T2K experiment, rely on millions of identical components. The SuperFGD consists of nearly 2 million plastic scintillator cubes (1 cm³), each containing three orthogonal 1.5 mm holes for 1.0 mm wavelength-shifting (WLS) fibers.

• The Challenge: During assembly, these cubes are stacked in a 3D grid. Even minor dimensional variations or hole misalignments can accumulate, preventing fiber insertion or causing fiber breakage due to stress.
• Precision Requirements: With 192 cubes aligned along the longest axis, the acceptable deviation per cube is calculated to be approximately 36 µm (assuming Gaussian distribution).
• Limitations of Existing Methods: The previous quality control method involved manual screening using stainless-steel rods. This process was:
  • Slow: ~16 seconds per cube (over 1 hour for a 15×15 batch).
  • Subjective: Relied on operator perception of rod insertion stress.
  • Inefficient: Required significant human labor and lacked quantitative reproducibility.

2. Methodology

The authors developed a two-stage semiautomatic system combining image analysis and robotics to measure cube dimensions, surface protrusions, and hole positions with high precision.

A. Optical and Mechanical System (Prototype)

• Imaging: Six 8-megapixel USB cameras equipped with 2× teleconverter lenses and ring-shaped LED lights provide uniform illumination. This setup achieves an effective resolution of ~15 µm/pixel.
• Mechanism: A motorized rotating stage holds eight platforms. A cube is placed on a platform, and the stage rotates 45° to capture three faces (Imaging Point 1), then another 45° to capture the remaining three faces (Imaging Point 2).
• Positioning: The platform uses gravity and textured walls to passively align the cube against three walls, ensuring precise positioning without active clamping.
• Sorting: A servo-controlled valve directs cubes into "Acceptable," "Defective," or "Reinspection" bins based on initial analysis.

B. Image Analysis Pipeline

The software uses OpenCV and custom algorithms to process images:

1. Preprocessing: Grayscale conversion, Gaussian filtering, and binarization to remove noise.
2. Contour & Shape Detection: Extraction of cube boundaries and hole perimeters. Hough transforms detect lines (cube edges) and circles (holes) to filter out minor surface irregularities.
3. Alignment Correction: Algorithms fit lines to cube edges to correct for systematic tilts in the camera or platform.
4. Illumination/Alignment Calibration: A linear fit correction is applied to account for lighting asymmetries and mechanical misalignments across the 6 cameras and 8 platforms, reducing measurement variance.
5. Protrusion Detection: Compares contours against ideal edges; bright pixels extending beyond the edge are quantified to detect surface defects.
6. Classification: Cubes are classified based on hole position ($x, y$), cube size ($L_x, L_y$), and protrusion index ($P_{bump}$).

C. Robotic Integration (Completed System)

To address the high rejection rate of the prototype, a 6-axis robotic arm (UFACTORY xArm 6) replaced the simple sorting valve.

• Function: The arm picks up the cube before it falls off the platform, preserving its orientation.
• Advanced Sorting: Instead of a binary pass/fail, the system classifies cubes into 48 groups based on the specific offset of their hole positions.
• Logic: Cubes with similar hole shifts are grouped together. When stacked in a row, their cumulative misalignment cancels out or remains within the 0.5 mm tolerance required for fiber insertion.
• Defect Handling: Cubes with severe size errors, protrusions, or Z-face hole deviations are rejected by the robot (allowed to fall into a defect bin).

3. Key Contributions

1. High-Precision Non-Contact Measurement: Developed a system achieving 10 µm measurement precision, significantly better than manual caliper measurements (which suffered from >30 µm variance due to surface softness and contact pressure).
2. Quantitative Quality Control: Replaced subjective manual screening with objective, reproducible image-based criteria.
3. Intelligent Sorting Strategy: Introduced a novel "grouping" methodology where cubes are sorted by their specific defect vector (hole shift) rather than simply rejected. This transforms a "defect" into a manageable parameter for assembly.
4. Robotic Automation: Successfully integrated a 6-axis robotic arm to maintain cube orientation and enable 48-way sorting, a feat impossible with simple gravity-fed valves.

4. Results

• Prototype Performance:
  • Speed: ~6 seconds per cube (3x faster than manual screening).
  • Precision: Standard deviation of hole position reduced from 23 µm to 10 µm after calibration corrections.
  • Reproducibility: Achieved >80% consistency with manual screening results when tested on a sample of 395 cubes (154 acceptable, 241 defective).
• Completed System Performance:
  • Throughput: ~15 seconds per cube (including robotic pickup).
  • Rejection Rate: Reduced the defect rate from ~20% (in the prototype) to 3.1% in a test run of ~6,500 cubes.
  • Grouping: Successfully categorized cubes into 48 groups where internal hole position variation was <500 µm, well within the tolerance for fiber insertion.
  • Reliability: The robotic arm operated stably with no significant failures (gripping or dropping) during testing.

5. Significance

This work presents a scalable solution for the quality control of future large-scale neutrino detectors (e.g., DUNE near detectors) which may require millions of components.

• Scalability: The system demonstrates that high-precision inspection can be automated, reducing human error and labor costs.
• Yield Optimization: By utilizing a "grouping" strategy rather than strict rejection, the system maximizes the usable yield of manufactured components, which is critical for cost-effective construction of massive detectors.
• Future Automation: The authors demonstrated that the robotic arm can also handle the initial loading of cubes, paving the way for a fully unattended, end-to-end automated inspection line using dual robotic arms.