CourtSweeper

A Vision-Based Tennis Ball Retriever with Dual-Roller Intake Mechanism.

Team #13: Yuwei ZHAO, Yilin ZHANG, Shijia LUO, Yingrui SUN

---

I. Introduction

This repository contains the software and hardware integration code for an autonomous tennis ball collection robot. The system utilizes a Jetson Orin NX to process visual and spatial data. It identifies tennis balls using a YOLO26 vision model and navigates the environment using a ROS2-based SLAM implementation. The physical ball collection is executed by a custom dual-motor friction wheel mechanism. A native Swift iOS application serves as the user interface for telemetry visualization and emergency control.

---

II. Demonstration

TODO

---

III. Project Structure

1. Mainbody

CourtSweeper/
├── doc/                        # Development Logs
│   ├── figures/
│   └── log/
│       └─ development_log.md
├── paper/                      # Report
│   ├── proj_proposal/
│   └── final_report/
├── ref/                        # Reference
├── ROS/                        # ROS-relative
│   ├── config/
│   └── map/
├── src/                        # Source code
│   ├── chassis/
│   │   └── ...
│   ├── distance/
│   │   └── ...
│   ├── lidar/
│   │   └── ...
│   ├── map/
│   │   └── ...
│   ├── roller/
│   │   └── ...
│   ├── vision/
│   │   └── ...
│   ├── env_config.py
│   └── main.py
├── User Interface/             # Swift App
├── LICENSE
└── README.md

2. Submodule

Source code part

Vision Module

vision/
├── mlmodel/
│   ├── engine/       # TensorRT model
│   ├── onnx/         # ONNX model
│   └── pt/           # Pt model
│
├── training/         # model training-related
│   ├── dataset/
│   ├── runs/
│   ├── weight/
│   ├── dataset_collect.py
│   └── model_training.py
├── cv_config.py      # cv-related env config
└── video_node.py     # load vision model

Chassis Module

chassis/
├── chassis_ctrl.py    # chassis mov ctrl
└── chassis_sub.py     # chassis info sub

Distance Module

distance/
└── distance_sub.py     # distance info sub

Lidar Module

lidar/
├── lidar_parse.py                    # RPLidar (test only)
└── x2_lidar_parse(deprecated).py     # YDLidar (test only)

Map Module

map/
├── map_generation.py    # slam mapping
├── map_navigation.py    # slam navigation
└── workflow.md

Roller Module

roller/
├── roller_drive.py
└── roller_setting.ino   # Arduino blink

User Interface part

User Interface (TODO)

User_Interface/
├── CourtSweeper.xcodeproj              # Xcode project config
├── CourtSweeper/
│   ├── App/
│   │   └── CourtSweeperApp.swift       # App entry point
│   │
│   ├── Models/
│   │   └── RobotCell.swift             # Data model
│   │
│   ├── ViewModels/
│   │   └── RobotControlViewModel.swift # State management hub
│   │
│   ├── Views/
│   │   ├── RobotControlView.swift      # Main control interface
│   │   └── Components/                 # Reusable UI components folder
│   │       ├── VideoStreamCard.swift   # Video player container
│   │       └── GridMatrixView.swift    # Grid matrix container
│   │
│   └── Assets.xcassets/                # Media asset catalog

---

IV. System Details

1. Hardware Architecture

• Chassis Platform: A DJI RoboMaster chassis provides mobility, camera input, infrared sensor and wireless communication.
• Edge Computing: A Jetson Orin NX serves as the primary computational platform. It operates headlessly using a display emulator and VNC.
• Spatial Mapping: An RPLidar A1M8 transmits 2D spatial data via serial communication. Due to chassis occlusion, the functional field of view is restricted to a 180-degree forward-facing arc.
• Actuation Control: An Arduino Uno acts as the lower-level microcontroller. It receives serial commands from the Jetson to drive the custom friction wheel intake mechanism using BTS7960 motor drivers and 755 motors.
• Power Distribution: Power is physically isolated across three independent circuits to maintain electrical stability - Jetson Orin NX: 12V 6800mAh battery; Intake Mechanism: 11.1V 3S 5200mAh LiPo battery; RoboMaster Chassis: Factory default battery.

2. Software Environment

• Operating System: The system runs on Ubuntu 22.04 with JetPack 6.2.1.
• Process Isolation: Vision and navigation stacks are isolated to prevent dependency conflicts.
• Vision Stack: The YOLO26 model executes within a dedicated Conda environment.
• Navigation Stack: ROS2 SLAM applications operate directly within the native system environment.
• Communication Bridge: A custom ROS2 node handles cross-environment data exchange. It interfaces with a modified SDK (Robomaster-SDK-Ultra) compatible with newer Python versions to manage velocity commands, odometry calculations, TF broadcasting, and image stream publishing.

The core dependencies are listed below:

Environment	Version
JetPack	6.2.1
Ubuntu	22.04 Jammy
CUDA	12.6.68
cuDNN	9.3.0.75
VPI	3.2.4
Vulkan	1.3.204
Python	3.10.12
OpenCV	4.11.0 with CUDA: YES
PyTorch	2.5.0a0+872d972e41.nv24.8 (torchvision 0.20.0a0+afc54f7)
TensorRT	10.3.0

---

V. Limitations and Future Work

1. Due to fabrication constraints, the mechanical stability of the custom dual-friction wheel structure requires further optimization.
2. SLAM mapping degrades in wide-open outdoor environments, leading to reduced navigation reliability.
3. The communication bandwidth of the RoboMaster chassis limits data throughput, constraining high-frequency data processing.
4. The current lidar configuration is restricted to a three-directional field of view. Future designs will implement an unoccluded 360-degree sensor placement.
5. The iOS application requires further feature development and user interface refinements.

---

⚠️ License: This project isn't open-source. See Details LICENSE.