A potential issue that was raised was the accuracy of a YOLO model, which tends to be used for applications where speed is more important than accuracy. We are currently looking into the possibility of more powerful object detection models, but are leaning towards a more recent version of YOLO, since it is easier to set up and run, and its wide use in past 18500 projects means that it is proven to work. Aside from the CV, a potential area of concern is the system to drop items into the bins. We’ve come up with an arrangement which should simplify the system as much as possible, with a linear gantry only requiring control to a stepper motor which will “select” a bin, and a servo to drop it in. In this way we avoid complex robot kinematics. If we end up switching to a single-stream system (more on that in the next paragraph), we will have only two categories, meaning the system could be further simplified.
Based on the feedback we received from our proposal presentation, we are considering changing our design to sorting into two bins, recyclable vs. trash, instead of four. Pittsburgh (and many other American cities) use single-stream recycling, which means all recyclable materials are collected together then sorted at the processing plant. Some places may collect sorted materials (ie. CMU has individual bins for different categories of recycling), but some places may simply collect our sorted materials and dump them together. Sorting into two bins will likely not have much impact on our computer vision system, which will still recognize and classify objects as we planned, the only difference being that classified objects will then be grouped based on whether or not they are recyclable, instead of the specific recyclable material (ie. water bottle → recyclable, rather than water bottle → plastic). There isn’t too much change on the hardware side, the system design will remain the same, only we have fewer categories. The biggest changes are in the mechanical design, primarily in part selection and physical size. With half as many categories, our bin can be physically smaller, and we can use a smaller gantry for item placing.
Since this change doesn’t impact the overall design of our project, we don’t anticipate our schedule to be impacted too much. This could have the potential to push back the hardware assembly part of the project, since we may be delayed in ordering the gantry, but we still have to go through the design process before starting assembly, so the delay should be small, if there is one.