Samueli School of Engineering

Due to growing waste concerns worldwide, we present WALTER, a novel autonomous robotic system capable of detecting and removing trash substances from a local environment. We approach this project through deploying a real-time trash detection model utilizing a YOLOv4-tiny backbone on a Xilinx FPGA with serial communication between a Raspberry Pi and Teensy micro-controller for decision making and motor control. WALTER is assembled with various 3D printed interfaces and collects trash using refashioned drawer rails as arms. The final trash detection model achieves a 37.0% average precision (AP) and the quantized model, running on a FPGA, operates at 5 frames per second (FPS) when running stereo vision object detection. Furthermore, WALTER is able to collect a diverse set of trash items in both indoor and outdoor environments, and is able to support up to a 2 kilogram payload.