The RoboMaster Design Team had no existing autonomous targeting capability; our robots relied entirely on manual control to identify and engage enemy armor plates in fast-paced competition :(Challenge
Solution!
V1 Python, Tensorflow, YOLOv3, Darknet, ROS, Azure ML
We started by building a custom dataset, capturing our own images of armor plates and manually annotating bounding boxes for each sample. Using this dataset for fine-tuning and testing, we built an object detection pipeline starting with early experimentation in TensorFlow's Object Detection API, before converging on Darknet and YOLOv3 as our production solution, with GPU-accelerated training on Azure Machine Learning for reproducible experiments and model versioning.
V2 Python, PyTorch, YOLOv5, ROS, NVIDIA Jetson Nano, GitLab
As AI Team Lead the following year, I directed a team of 15 engineers through a full overhaul, revisiting and expanding the dataset before migrating to YOLOv5 deployed via ROS on NVIDIA Jetson Nano robots for efficient on-device inference.