ROBOTIC HAND SORTING OF VLSI DEVICES - VLSI Pedia Online VLSI Courses & Training

ROBOTIC HAND SORTING OF VLSI DEVICES

(A practical engineering guide for automation teams, integrators, and fab engineers)

Executive summary

Robotic hand sorting of VLSI devices replaces manual pick-and-place operations for die, trays, tape-and-reel, waffle packs, and small packages. A successful system combines:

high-resolution vision for part detection & orientation,
mechanical end-effectors (vacuum / soft / micro-grippers) tuned for fragile parts,
ESD/cleanroom compliant materials and procedures, and
a control stack (real-time motion + vision + PLC/MES) for deterministic throughput and traceability.

Goals: maximize throughput, minimize damage and mis-sorts, ensure traceability (serials/lot), and maintain cleanliness & ESD safety.

1. Use cases & scope

Common scenarios:

Sorting bare dice from singulation output (die sorting).
Picking packaged chips from trays/waffle packs to tape or cassette.
Rejecting defective parts after test (in-line sorting).
Re-diverting items by part number, grade, or blind binning for rework.

Parts handled: small SOT/QFN/TSOP packages, bare die (0.5–10 mm), bumped die, leaded packages — each has distinct handling requirements.

2. System architecture (high level)

Material feed / input — conveyor, vibratory feeder, tray/vial feeder, singulation line.
Pre-inspection station — coarse vision to check orientation, contamination.
Precision vision & pose estimation — high resolution camera(s), telecentric lens, structured light if needed.
Robotic manipulator — 3–6 axis robot or SCARA for fast XY moves + rotation; sometimes 6-axis for flexibility.
End-effector / robotic hand — vacuum, soft pneumatic gripper, micro-mechanical gripper, or hybrid.
Placement / output — destination trays, tape & reel, reject bin, or downstream process.
Control layer — real-time motion controller (robot controller), vision processor (GPU/FPGA/embedded PC), PLC / MES connect.
Test / verification — post-place vision check, electrical probing if required.
Traceability — barcode/RFID reader, MES logging, serial association.

Diagram (conceptual):

3. Key hardware choices

Robot type

SCARA — excellent for high-speed XY motion, limited Z/rotation flexibility. Good for planar trays/tape.
6-axis articulated — high flexibility for complex orientations and 3D access but usually slower. Best for die handling with non-planar needs.
Cartesian gantry — high repeatability and throughput for linear conveyor lines.

Choose based on workspace, payload (very low for dice), speed, and repeatability (<±20 µm often required).

End-effector (gripper) options

Vacuum nozzle — common for flat, smooth surfaces. Requires vacuum cup size matched to device; risk: chip pick-up failure on porous surfaces.
Soft pneumatic gripper (silicone fingers) — better compliance, distributes force, reduces chipping.
Micro-mechanical gripper — tiny jaws for edge gripping (good for bumped die).
Electrostatic chuck / die ejector — useful for wafer/die pick where electrostatic hold aids precision.
Hybrid — vacuum + mechanical clamp for redundancy.

Gripper design must minimize lateral shear, control approach speed, and limit normal force to avoid cracking. Materials must be non-abrading and ESD-safe.

Vision & sensing

Cameras: 5–20 MP area sensor or line scan depending on throughput. Telecentric lenses avoid perspective distortion for precision.
Illumination: ring lights + coaxial for shiny leads, backlight for silhouette/edge detection, structured light for 3D pose if needed.
Proximity/force sensors: force / tactile sensors for touch verification and safe placement.
ESD sensors: monitor wrist strap, ionizers, and operator interlocks.

Processing & control hardware

Vision processor: GPU (for neural pipelines), FPGA (low latency), or embedded CPU for classical CV.
Motion controller / PLC: deterministic robot control, safety IO, and integration to MES.
Real-time IO: EtherCAT/Profinet for low latency.

4. Software stack & algorithms

4.1 Vision pipeline

Acquisition — synchronized capture with motion.
Preprocessing — denoise, contrast, normalize.
Segmentation — detect parts using classical (thresholding + morphology) or trained CNN (faster R-CNN, YOLO) for multiple part types and occlusion.
Pose estimation — subpixel corner detection or neural pose regression to get X,Y,θ (and Z if necessary). Telecentric optics simplifies math.
Defect detection — surface scratches, contamination, peeling using an anomaly detection network.
Barcode / OCR / Mark read — for lot traceability or die ID.

Recommended approach: use a hybrid: fast classical for obvious cases (contrast OK), ML model for robust detection in variable lighting/part types.

4.2 Motion & grasp planning

Convert vision pose to robot coordinates using calibrated extrinsics.
Plan approach vector to minimize shear — vertical approach followed by slight tilt for edge grip if needed.
Use force feedback: retract if contact force exceeds threshold.
Add retry logic with small perturbation if initial pick fails.

4.3 Sorting logic & rules engine

Integrate MES rules: part grade → bin A, B, rework.
Real-time decisions: based on vision confidence, electrical test result, or operator overrides.
Logging each pick/place with timestamp, camera image, and robot state for traceability.

4.4 Sample pseudocode (pick-place loop)

loop:

  frame = capture()

  detections = detect_parts(frame)

  for part in detections:

    pose = estimate_pose(part)

    if pose.confidence < CONF_THRESH:

      log("low_confidence", part)

      continue

    plan = plan_motion(pose, dest=select_destination(part))

    execute_motion(plan)

    if not verify_pick():

      retry_count += 1

      if retry_count <= MAX_RETRY:

        adjust_pose = perturb(pose)

        plan = plan_motion(adjust_pose, dest)

        execute_motion(plan)

      else:

        divert_to_reject()

    else:

      place_and_verify()

      log_success()

5. ESD, cleanliness & mechanical care

ESD: All end-effector materials conductive/ESD-safe; active grounding of robot wrist; ionizers at pick zones; continuous ground monitoring.
Particle control: choose low-particle materials (silicone wipers that don’t shed), regular cleaning schedule, laminar flow hood if in cleanroom.
Humidity control: some die are humidity-sensitive — maintain specified RH.
Force limits & motion profiles: ramped speeds, soft landings, constant suction parameters to avoid crack.
Handling adhesives: avoid adhesive residue or tapes that leave tack on devices.

6. Throughput & performance planning (with worked examples)

Throughput depends on cycle time (full pick→place→return) and parallelism.

Throughput formula:

$(s)\text{throughput (parts/hour)} = \frac{3600}{\text{cycle time (s)}}$

Example 1 — single robot, cycle time 0.5 s:

Calculate: 3600 ÷ 0.5 = 7200 parts/hour.

Example 2 — requirement: 10,000 parts/hour. Required cycle time per pick:

cycle time = 3600 ÷ 10000 = 0.36 s.
So each pick-place must be ≤ 0.36 s; evaluate if robot, vision, gripping and movement can meet that.

Multi-robot or multi-gripper: if two identical parallel robots share workload, required cycle time per robot doubles (i.e., each robot handles half). Example: two robots each with 0.72 s cycle → combined 7200*2 = 14400/hr theoretical.

Important practical points:

Vision & motion overlap: pipeline acquisition while robot moves to maximize efficiency.
Add margins for retries and inspections: design for 80–90% nominal to meet realistic targets.

7. Quality, testing & verification

Key KPIs

Yield of undamaged parts (%) — target >99.9% for packaged parts, depending on spec.
Pick success rate — percent picks without retries.
Sort accuracy — correct destination assignments.
Cycle time — average and 95th percentile.
Traceability completeness — photos and logs per item.

Test plan

Baseline tests: repeated pick/place of dummy parts to measure repeatability and pick force tuning.
Stress tests: long-run with different part orientations to detect wear/contamination.
ESD tests: verify wrist/ionizer performance.
Contamination & particle counts: periodic examination, especially for bare die handling.

Post-implementation verification

Integrate electrical spot checks on random samples to ensure no latent damage (e.g., parametric tests after sorting).

8. Traceability & integration with MES

Each device pick logged with: timestamp, lotID, image, pose, robot ID, gripper type, operator.
Use barcodes / OCR or 2D code readers to associate lot/channel.
Provide API for MES to query counts, reject rates, and images for audit.

9. Failure modes & mitigations

Pick failure (no suction): check vacuum leak, surface contamination, adjust vacuum level or suction cup size.
Mis-grip / chip slip during transport: reduce transfer speed, add soft clamp.
ESD event causing device damage: add continuous ground monitoring, ESD event logging, fast interlock.
Vision false positive / missed detection: augment dataset, improve illumination, use multi-view cameras.
Debris accumulation on nozzle: implement automated nozzle cleaning cycle and schedule.
Throughput bottleneck at vision stage: parallelize vision or use simpler classical heuristics where possible.

10. Implementation roadmap (step-by-step)

Requirements & constraints — parts types, throughput, floor space, cleanroom class, ESD class, MES interface.
Feasibility & risk analysis — handle fragility, surface finish, and orientation variability.
Select robotics & gripper tech — prototype multiple grippers on sample parts.
Design vision pipeline — collect training images under production lighting; prototype with offline processing.
Integration & control — choose motion controller, I/O, and real-time buses.
Prototype cell — build single-robot cell, iterate on picks & tuning.
Test & qualification — run acceptance tests: throughput, yield, electrical checks.
Scale & redundancy — add parallel cells and automated material handling for continuous flow.
Operator training & SOPs — cleaning, ESD checks, failure response.
Maintenance & continuous improvement — scheduled wear replacement, vision retraining with new part lots.

11. Practical BOM checklist (select items)

Industrial robot (SCARA or articulated) with high repeatability.
Precision linear stages or gantry for conveyor integration.
High-resolution camera(s) + telecentric lenses.
Structured/LED illumination modules.
Vacuum pumps, valves, and vacuum cups; or pneumatic soft gripper hardware.
Force sensor or tactile sensor module.
Real-time motion controller (with EtherCAT support).
Vision processor (GPU or FPGA); industrial PC.
PLC for safety I/O, light curtain, interlocks.
Ionizers, ESD wrist & grounding monitors.
MES connectivity software, database server, barcode/RFID reader.

12. Cost vs benefit considerations

CapEx: robots, vision, cleanroom interface, MES integration — significant but amortized over throughput & labor savings.
OpEx: maintenance, vacuum pumps, nozzle wear parts, vision model retraining.
Benefits: higher consistent throughput, reduced human error, better traceability, reduced contamination risk, 24/7 operation.

Return on investment depends on volume, labor cost, yield improvement, and reduced rework.

13. Regulatory & safety considerations

Safety interlocks, emergency stop, light curtains, robot cage where appropriate.
For cleanroom environments, ensure robot/cable routing meets cleanroom standards (materials, lubricants).
ESD compliance to relevant industry standard (keep within fab class ESD guidelines).

14. Example case study (hypothetical)

Objective: retrofit a die singulation line to sort good die into waffle packs at 6,000 parts/hour.

Required cycle time = 3600 / 6000 = 0.6 s per pick.
Chosen robot: high-speed SCARA capable of 0.35 s average cycle for planned motions (vision and motion overlapped to meet 0.6 s).
End-effector: hybrid vacuum + soft clamp for bumped/BGA-like die.
Vision: 5MP telecentric camera with backlight and top diffuse ring; CNN for defect scoring.
Throughput achieved after pipeline optimization: 6,300 parts/hour with <0.1% mechanical damage rate.

15. Final recommendations (quick list)

Match end-effector to the part surface type and fragile features.
Prioritize telecentric optics for subpixel pose accuracy.
Pipeline vision with overlapping robot motion to maximize throughput.
Build ESD & cleanroom compliance from day one — retrofitting is costly.
Use data logging / images for continuous QA and ML retraining.
Prototype early with representative parts; iterate gripper and vision before full integration.

VLSI Expert India: Dr. Pallavi Agrawal, Ph.D., M.Tech, B.Tech (MANIT Bhopal) – Electronics and Telecommunications Engineering