Case Study: RoboSub Team SONIA achieves a top-3 finish by accelerating dataset creation and improving model performance with Labelbox

RoboSub - Customer Case Study

RoboSub’s SONIA team—one of the long-running student entrants in the RoboSub competition—faced a last‑minute rule change three weeks before the event that rendered their initial visual training data unusable for the Cash In task. To recover under a tight timeline they used Labelbox’s free academic license and cloud labeling tools to rapidly create and manage new training datasets for their AUV’s perception models.

Using Labelbox, the team added roughly 1,500 new labels to reach about 10,000–11,000 images per object and leveraged the Labelbox GraphQL API and distributed labeling features to scale annotation quickly. Labeled data were exported, converted to TFRecords, and trained with the TensorFlow Object Detection API (about 5 hours/model on an NVIDIA P100), while Apache Airflow automated the ML pipeline. The result: SONIA secured a top‑3 finish in the competition and established an automated, repeatable workflow for future RoboSub events, demonstrating measurable dataset growth and competitive performance enabled by Labelbox.

RoboSub