Full technical video12:29 · seven-person team
Preparing seekable video…
Jump to a chapter
Team 8 technical review
Open video ↗The full prototype walkthrough covering motivation, mechanics, electronics, software, AI/ML, testing, and integration.
Drosophila Automation Platform
AI-Powered Robotic Platform for Genetic Screening:
Built an autonomous system that identifies, sorts, and evaluates fruit flies for genetics research. Integrated computer vision, embedded controls, and custom hardware to achieve ~97% classification accuracy, earning Clemson’s People’s Choice Award with a cross-functional group of 7 people.
7person senior-design team
3camera system architecture
>97%sexing accuracy · 45/46
1People’s Choice Award
01
Project record
PROJECT DOCUMENTATION
The full technical walkthrough, presentation, and final report.
Together, the video, presentation, and final report document the system architecture, mechanical and electrical design, software stack, integration tests, and results.
Team-authored final report25 pages · Spring 2026
Fly Sorter and Loader
The latest team-authored design record, including requirements, subsystem decisions, implementation, testing, integration architecture, and results.
Team technical presentation16 slides · Spring 2026
Automated Drosophila Sorting System
The team’s project-review deck covers the workflow, 3D-printed hardware, PCB and wiring, GUI, computer vision, testing, reported results, and next steps.
The team, prototype, and showcase.
The project culminated in a physical prototype, technical poster, and People’s Choice recognition at Clemson’s Senior Design II Showcase.
02 · Research impact
Why automate this workflow?
The platform links preparation, classification, behavioral measurement, and research outputs in one traceable path.
Hover, focus, or select a stage.
Turn a mixed vial into organized experimental groups.
A typical source vial held roughly 40–50 flies that had to be sedated, separated, counted, and transferred into clean vials.
The channel constrains the group, the overhead camera localizes candidates, and the gantry with vacuum pickup moves flies through a controlled path.
A coordinate-linked transfer workflow is designed to reduce repeated handling and make each motion step measurable, reviewable, and easier to improve.
SYSTEM MAP
Computer vision and automation pipeline.
Camera views
Three calibrated imaging stations
Computer vision
Channel detection + millimeter mapping
Decision + control
Hybrid ML, gantry, vacuum handling
Assay + outputs
Tracking, video, CSV, summaries, reports
Embedded + mechanical layer
Raspberry Pi 5 supervisory compute · single-axis gantry · vacuum pickup · vibration-assisted handling · motors, sensors, and custom electronics
03
Inside the system
A coordinated physical and computational workflow.
Three camera views, Raspberry Pi compute, computer vision, machine learning, motion, vacuum handling, behavioral tracking, and an operator interface were integrated into one coordinated prototype workflow.
Hover, focus, or tap a stage to inspect its hardware, specifications, and project visuals.
- 01Load
- 02Localize
- 03Pick + image
- 04Classify
- 05Route + stage assay
- 06Record + measure
- 07Review
STEP 01 · Input handling
Load
Sedated flies enter a CO₂-supported channel with geometry designed to separate them for pickup.
- Hardware
- 3D-printed channel + under-channel CO₂ manifold
- Input
- Sedated Drosophila loaded into the channel
- Geometry
- Center ridges/dimples, lateral vent holes, ~25-fly design capacity
- Next stage
- Whole-channel overhead image
AI / ML DECISION PATH
Learned appearance + physical morphology
>97%sexing accuracy · YOLO test · 45/46
YOLO supplied a learned class prediction and confidence. An image-segmentation branch supplied 28 interpretable morphology features. A Random Forest fused those with YOLO class and confidence, while the wider workflow retained an uncertain route.
The >97% sexing result comes from a focused YOLO test (45/46). The separate hybrid model measured 94.2% test accuracy on an 80/20 split of a 1,000-image mixed dataset.
>97% sexing accuracy · open ↗
28 morphology features · open ↗
94.2% hybrid test · open ↗
Calibration view ↗01Computer vision
Channel computer vision
The first perception stage turns a live camera frame into physical pickup coordinates—not just bounding boxes.
- Registers live frames to an empty-channel reference using an ECC transform.
- Applies black-hat morphology, ROI masking, thresholding, cleanup, and connected-component filtering.
- Projects accepted centroids onto the calibrated channel axis and converts pixels to millimeters.
Output
Fly count, ordered positions, annotated image, binary mask, and structured detection data.
Prediction review grid ↗02AI / ML
AI/ML classification
The classifier combines learned visual features with interpretable morphology to produce a reviewable decision.
- Fine-tunes a lightweight YOLO model on labeled male and female fly images using transfer learning and controlled augmentation.
- Development included label, class-balance, duplicate-image, and background-bias audits alongside morphology-based segmentation.
- Extracts 28 morphology features, then adds YOLO class and confidence to form the 30-feature Random Forest input.
Output
A reviewable class decision tied to a specific imaging setup and evaluation protocol.
Calibrated assay tracking ↗03Assay analysis
Climbing-assay analysis
The assay stage translates calibrated multi-vial video into track-level and vial-level behavioral measurements.
- Stores a profile with camera, transform, background, detector settings, and vial-specific regions and baselines.
- Runs background subtraction, dark-object enhancement, morphology, connected components, and within-vial matching frame by frame.
- Separates recording from processing after Raspberry Pi and camera constraints made simultaneous analysis unreliable.
Output
Position, distance, speed, velocity, threshold crossings, annotated video, and per-vial summaries.
Results + report viewer ↗04Data + interface
Data + researcher GUI
The software preserves intermediate evidence so operators can inspect how each result was produced.
- Exposes calibration, background capture, transforms, device status, recording, processing, playback, and export.
- Keeps calibration and profile settings alongside raw and processed run artifacts for review and debugging.
- Organizes frame-, track-, vial-, and run-level outputs into researcher-readable artifacts.
Output
CSV, SQLite, JSON metadata, annotated media, summaries, and PDF reports.
Assembled physical prototype ↗05Embedded robotics
Embedded + mechatronics
A Raspberry Pi 5 supervised the cameras and machine state while custom electronics connected perception to physical handling.
- Coordinates the Logitech channel camera, microscope classification camera, and fixed climbing-assay camera.
- Controls the belt-driven gantry, stepper driver, vacuum pump and nozzle, vibration motor, limit switches, and lighting.
- Uses a modular 3D-printed table and custom PCB to bring mechanical, electrical, and software subsystems together.
Output
A physical prototype that connected vision performance to real-world motion, alignment, airflow, and handling reliability.
TECHNICAL ARTIFACTS
From PCB design to calibrated vision.
Explore the electrical, mechanical, computer-vision, and data interfaces that connected the prototype end to end.
Custom control PCB
The custom KiCad board organizes Raspberry Pi GPIO, motor-driver interfaces, actuator power and control, and limit-switch inputs.
Project archive · PCB design
Core PCB schematic
The schematic maps Raspberry Pi GPIO, TMC2208 stepper interfaces, vacuum and vibration motor drivers, power, and limit-switch connections.
Project archive · core PCB schematic
Mechanical CAD assembly
The exploded assembly shows the modular three-section table—channel, sexing and imaging station, and climbing-assay hardware—with independently replaceable printed parts.
Final report · pp. 6–8
3D-printed channel + linear gantry
This prototype iteration shows the printed channel and modular table section beside the belt-driven linear gantry hardware.
Final report · pp. 8–9Channel calibration workflow
The calibration view exposes the channel-localization pipeline: reference capture, region setup, image processing, and physical position mapping.
Final report · pp. 9–10Assay-analysis tracking
The assay workspace converts calibrated vial video into inspectable tracks and behavioral measurements.
Final report · pp. 12–13Data + report review
The data and GUI layer brings configuration, assay outputs, annotated media, and researcher-readable reports into one review workflow.
Final report · p. 17ML prediction review
The prediction grid enabled image-by-image review of model output and recurring failure cases.
Final report · pp. 14–16PCB DESIGN EXPLORER
Trace the control board from schematic to 3D assembly.
The team designed a custom KiCad PCB to simplify system wiring, organizing Raspberry Pi 5 GPIO, stepper- and DC-motor driver interfaces, 12 V power input, actuator terminals, and limit-switch connections.
Open full size ↗Assembly
Populated-board overview
The assembled view shows the Pi header, driver sockets, power terminals, actuator connections, limit-switch headers, and board-level routing in one frame.
- Compute
- Raspberry Pi 5 GPIO header
- Motion
- Stepper + DC-motor driver interfaces
- I/O
- Actuator terminals + limit switches
THREE-CAMERA ARCHITECTURE
Each view had one job.
Hover, focus, or tap a camera to locate its station on the prototype.
VACUUM + MOTION TESTS
Watch the pickup head move through the channel workspace.
Two short project clips show the gantry-vacuum mechanism at different scales: the full travel path and a close-up at the illuminated pickup area.
OFFLINE ASSAY ANALYSIS
Watch the tracker preserve each fly’s path inside its vial.
This 10-second project output shows calibrated vial regions, per-fly boxes and identifiers, coordinates, a threshold line, and frame-level counts.
04
Bottlenecks + iterations
The prototype improved where software met physics.
The most useful lessons came from isolating a failure, changing one physical or computational assumption, and testing the stage again.
Hover, focus, or tap an engineering challenge to inspect the design response and evidence.
01Channel geometry
Make each fly reachable before asking vision to find it.
- Bottleneck
- A single-axis nozzle could only pick flies that were separated and sitting within a narrow reachable height range.
- Design response
- Successive printed channels explored center ridges and dimples for isolation, side vents tied to the CO₂ manifold, and brush-assisted distribution.
- What the evidence says
- The final channel was designed for roughly 25 flies. The postmortem still identified channel length and clustered flies as limits for a future build.
A correct prediction still fails the experiment when the channel, nozzle, camera, and motion system disagree on where the fly is.
Core integration lesson
05
Evaluation results
Measured at the subsystem and prototype levels.
The project evaluated sex classification, hybrid fusion, and the physical movement sequence under separate protocols.
Hover, focus, or tap a result to see how it was evaluated.
ACTIVE RESULT>97%
Sexing accuracy
- Evaluation
- YOLO classification
- Scope
- 45 / 46 test images classified correctly
- How to read it
- A focused diagnostic on a test sample.
TEAM RECOGNITION
People’s Choice Award
Clemson Senior Design II Showcase · April 2026
06
Limits and next iteration
WHAT THIS WAS
What worked—and what the next iteration needs.
01Throughput+
Approximately four flies per minute was a design target, not a verified sustained end-to-end result.
02Dataset shift+
Lighting, focus, angle, and camera setup could move performance away from results measured on saved-image evaluations.
03Physical reliability+
A correct coordinate could still fail because of nozzle height, alignment, airflow, channel geometry, or motion error.
04Assay artifacts+
Vial edges, overlapping flies, lighting changes, and camera constraints could create false detections or broken tracks.
05Deployment+
The prototype was designed for senior-design validation, not routine laboratory deployment.
06Next build+
Expand the evaluation dataset, label failures by stage, and run complete cycles under a fixed protocol.
PROJECT CREDIT
Built with a team, for a research sponsor.
Seven-person Clemson University senior-design team · Institute for Human Genetics sponsor · Spring 2026.