About the Competition

Our robot was based on a turtlebot3 customized, for the navigation software we used ROS noetic meanwhile thar for the detection part we used YOLO with specially trained weights for the competition.

The hardware modification included: 

• 2 IMX219 cameras.

• 2 Jetson boards (One for ROS & one for computer vision).

• Lower the Lidar position.

• Custom PCB for LEDs indicators and Buck-Boost for power distribution. 

To decrease the programming and debugging time an exact copy of the competition scenario was created on Gazebo, including apple tree models, our robot and obstacles.

For the navigation part, as our robot have odometry from the encoder, an IMU and a Lidar sensor AMCL was used for localization. Meanwhile, TEB and DWA were tested as the local planner in the simulation and in the real scenario, being the first one chose for our robot.

The Dynamic Window Approach (DWA) offers real-time performance and effective obstacle avoidance for differential drive robots due to its simple implementation, but it struggles with long-term path planning, suboptimal paths, and parameter sensitivity, particularly in dynamic environments. In contrast, the Timed Elastic Band (TEB) algorithm excels in producing optimal, smooth paths and handling dynamic obstacles by optimizing trajectories in real-time, making it flexible for various robot types. However, TEB is more computationally intensive, complex to implement, and requires careful parameter tuning, which can challenge real-time performance in resource-constrained systems.

For creating the map Karto SLAM was used because it combines LIDAR and IMU to enhance accuracy and adaptability to different scenarios. Besides the advanced optimization algorithms used by Karto SLAM ensure precise and reliable map generation.

To detect apples on a tree using YOLOX with a custom dataset, images of apples were collected and annotated under various conditions, ensuring diverse lighting, angles, and occlusions. The dataset was prepared in the YOLO format with labeled bounding boxes. The YOLOX model was then trained using this dataset by setting up the training environment with the necessary dependencies, configuring the model parameters, and running the training script.  The advantages of using YOLOX include its state-of-the-art performance in object detection, efficiency in handling large datasets, real-time processing capabilities, and robustness in detecting objects in complex environments, making it particularly well-suited for agricultural applications like apple detection.

After a hard day of competition, we were thrilled to secure second place and a prize of 3,000,000 Won. The day was filled with intense challenges and fierce competitors, making our achievement even more rewarding. Every team member gave their best effort, and our dedication and teamwork truly paid off. The experience was not only a test of our skills and endurance but also a demonstration of our ability to work together under pressure. The 3,000,000 Won prize is a testament to our hard work and commitment, and it serves as a motivation for us to strive for even greater success in future competitions.

• GitHub