2026ArchivedExperiments
Tennis-Ball Robot
tennis-ball-robot
A graduation project: a robot that collects tennis balls — architecture first, hardware second.
- Role
- Design & development, solo
- Year
- 2026
- Status
- Archived
- Stack
- ROSEmbedded CPython3D Printing
- HardwareL1 · L1Two-wheel differential chassis, roller pickup mechanism, 3D-printed structure.
- BehaviourL2 · L2“Find the ball, approach, pick up” as composable skills.
- TransportL3 · L3State and commands between the embedded controller and the on-board computer.
- ApplicationL4 · L4Decides what happens next and sequences the task.
- ViewL5 · L5Exposes state for debugging and demos.
Graduation projects are usually judged on a demo. I treated mine as an architecture exercise first: collecting tennis balls on a court is a system problem long before it becomes a motor problem.
The problem
Balls scatter across the court after every drill, and picking them up by hand is slow and hard on your back. Commercial collectors exist, but they are expensive, heavy, and assume a flat, well-lit court. My brief was more specific: design a small robot that can collect balls on an ordinary campus court, and make it buildable by one student in one semester.
What I built
Architecture first, hardware second. Five layers: hardware, behaviour abstraction, transport, application and view.
- Hardware: a two-wheel differential chassis with a roller pickup mechanism, 3D-printed around off-the-shelf motors.
- Behaviour abstraction: turns “find a ball”, “approach” and “pick up” into composable skills.
- Transport: carries state and commands between the embedded controller and the on-board computer.
- Application decides what to do; the view layer exposes state for debugging and demos.
The point of the layers is substitutability: a behaviour should not know whether it runs against a simulated chassis or the real one, and the view should not care how the robot is built.
Architecture notes
The plan was scheduled in 16 weeks: research and requirements → chassis and mechanism → control and ROS integration → vision → integration and testing. That schedule forced an early question: if I could only do one thing well, which one? My answer was the mechanism — the roller’s geometry decides how fast the robot can drive, and no amount of control code fixes a bad pickup path.
Vision uses a camera to detect balls and hands a target to the behaviour layer. It is the easiest part to overestimate, so it stayed last.
What I learned
First, mechanism is a software constraint; choosing the wrong hardware leaves software with very little room. Second, a layer diagram is only useful if it changes who owns a decision, otherwise it is decoration. Third, scope honesty matters: what I finished is a defensible system design with an honest prototype budget, not a polished product.
The thesis is archived and the repository is not open source. Looking back, the project taught me less about ROS than about thinking the system through before touching hardware.