LitterBot AI
EfficientNetV2 CNN that sorts a photo of waste into 9 recycling categories and tells you which bin it belongs in, built in Litterbot design, achieved DSWA innovation award at Diamond Challenge finalists.
Image classifier that takes a photo of a single piece of waste and predicts which of 9 categories it belongs to: cardboard, food organics, glass, metal, miscellaneous trash, paper, plastic, textile trash, or vegetation.
Built on EfficientNetV2-B0 with transfer learning. The backbone was pretrained on ImageNet, then fine-tuned on the RealWaste dataset (about 4,800 photos taken under realistic landfill lighting, not studio shots). Heavy on-the-fly augmentation (rotation, shifts, brightness jitter, horizontal flips) plus class-weighted loss handled the dataset imbalance. The fine-tuned model reaches about 87.6% validation accuracy.
The image classifier was built for Litterbot, a robot that autonomously navigates, picks up, and sorts pieces of trash. The proposed design gives Litterbot cameras that use a YOLO algorithm to scan its surroundings and identify pieces of trash. Litterbot then moves to the trash and uses a separate EfficientNetV2-B0 CNN to classify what type of trash it is. Litterbot qualified for the Diamond Challenge Finalists Limitless World Summit, where it won the Delaware Solid Waste Authority (DSWA) Waste and Recycling Innovation award.
The source code, training notes, and model files are available in the GitHub repo .