Building a Plant-Cloning Robot Prototype
A computer-vision robotics prototype for automating plant micropropagation, built from concept to working demonstration in several months using commodity robotics hardware, custom machine-learning models, embedded control, electronics, and rapid prototyping.
Video transcript summary
YouTube auto-captions available; summary on this page supplements the demo video.
Shows the micropropagation plant-cloning robot identifying and working with plant tissue using computer vision: commodity arm mechanics, camera-based segmentation, and automated cutting and placement workflow.
My role
Technical Lead. Full-system technical leadership across architecture, computer vision, machine learning, software, robotics, embedded firmware and electronics, mechanical prototyping, testing, and public demo activity. I worked across the complete system rather than a single subsystem.
Overview
At an early-stage agricultural robotics startup, I led technical development of an experimental robotic system for automating parts of the plant micropropagation workflow.
The project combined computer vision, machine learning, robotics, embedded systems, electronics, and mechanical prototyping. The goal was to prove that delicate plant-processing operations could be partially automated using relatively inexpensive hardware and a custom perception and control stack.
The prototype reached a working demonstration within several months.
My work included:
- overall system architecture and technical direction
- computer vision and custom image-segmentation models
- machine-learning training and data preparation
- Python and C/C++ software
- robotic control and inverse kinematics
- ROS-based integration
- embedded firmware and STM32-based electronics
- sensors, schematics, and hardware integration
- mechanical prototyping and 3D-printed components
- calibration, testing, and iterative prototype development
I also presented the system publicly as part of the project’s demo and startup activities.
The engineering problem
Plant micropropagation involves repetitive manual operations that require visual recognition, careful positioning, and physical manipulation of delicate biological material.
The prototype explored whether those tasks could be automated with a relatively low-cost robotic system.
Conceptually, the system followed this flow:
- Camera
- Image acquisition
- Custom segmentation model
- Plant/tissue localization
- Robot motion planning
- Embedded motor control
- Physical manipulation
The challenge was not simply recognizing plant tissue in an image.
The perception system had to connect to physical coordinates, robotic motion, embedded hardware, and a mechanism capable of acting on the result.
Computer vision and machine learning
A major part of my work was developing a custom segmentation neural network for identifying relevant plant structures from camera images.
That involved the complete ML workflow:
- collecting and preparing image data
- labeling training samples
- training and evaluating segmentation models
- processing model output
- converting image-space results into information usable by the robotic system
- iterating on the model together with the physical prototype
This was practical machine learning rather than an isolated research model: the output ultimately had to drive a real machine.
Robotics and embedded systems
The robot combined higher-level perception and planning with embedded control.
Technologies and areas involved included:
Python · C/C++ · TensorFlow · ROS · computer vision · CNNs · image segmentation · inverse kinematics · Google Edge TPU · STM32 · electronics · schematics · rapid prototyping
Commodity and reusable hardware were deliberately used wherever possible.
The objective was to validate the automation concept quickly rather than spend the early prototype phase designing every mechanical component from scratch.
That approach allowed the team to iterate rapidly between:
- Vision
- Software
- Firmware
- Mechanics
- Experiment
- Repeat
Why this project matters to me
This project is a good example of the kind of engineering I enjoy most.
There was no single technology that solved the problem.
A useful result required machine learning, software architecture, robotics, embedded control, electronics, mechanics, experimentation, and a lot of integration work.
It also reinforced a principle that has stayed with me: for an early-stage technical product, the fastest route to useful knowledge is often to combine existing components with custom engineering where it actually matters.
Build the smallest system that can test the hypothesis, measure what happens, and iterate from there.
Outcome
Within several months, the project progressed from an engineering concept to a working robotic prototype and public demonstration.
The underlying computer-vision and electronic architecture later remained valuable enough to be considered for reuse in other agricultural robotics applications.
There is no public source-code repository or published model output for the project, so the strongest public evidence is the working prototype and recorded presentations.