Karen:
This week I was able to completely finish writing and debugging the script for basic movement. But we realized that the raspberry pi did not come with a SD card. So I was not able to set up the pi and upload the code onto it. Because of this, I put in an order for an SD card and reader- hopefully this will come in by Monday so that we can get that part moving. With this delay, I started writing wrapper functions to control the Create2. This way, instead of having to deal with calculating the angle of rotation so that it is scaled or matching speed to distance, I can call a function to move a certain angle or amount of meters. If time permits, I would like to add to these wrapper functions so that interaction with the Create2 will not be highly involved every time a command is sent. My next goals for the next week are to get the pi set up, once the SD card arrives and then get the bot moving. I also would like to find easily connectable sensors to the pi that will help with detection of obstacles and order them. We also have our Design review presentation this week, so hopefully we will be able to add to our design after we receive feedback.
In terms of our initial schedule, we are slightly behind. I had hoped to already have the roomba moving by now. However, I do not think this is as large of an issue because most of the delay is because of missing parts. All the set up is ready, it just needs to be uploaded. If however, this ends up being an issue, I will just have to spend more time outside of what has already been allocated for the project and resolve this.
Joseph:
This week Manini and I finalized the model that we are using. After running more tests and research into Yolo, we decided that the drop in accuracy was not worth switching away from the RCNN model. Manini and I have begun the implementation process of the model on Pytorch and are looking into various methods of biasing it towards false positives as we believe it is important for our robot to never miss any humans. My goal next week is to help finish the initial implementation of the model so we can load a pre-trained model in and see the general results of how the model works. Afterwards, I will move on to working on the ability to wirelessly send images between the pi and my computer.
As for our initial schedule, we are slightly behind as we would have liked to finish the model by this week and started retraining/experimenting with biasing next week. However, we do have an extra week of slack built into our schedule just for the deep learning implementation so this should not affect our timeline too much.
Manini
This week Joseph and I finalized the deep learning model we will be using for human detection. After some more researching, we came to the conclusion to stick with Faster RCNN since it produces a better accuracy score and since we are not trying to achieve real time detection or trying to run our model on a pi (inference will be done on a server). Joseph and I also began the implementation of the Faster RCNN model using PyTorch. According to our initial schedule we are slightly behind as we hoped to have a basic implementation of the model completed by the end of this week. However, with the time spent on finalizing the model we were not able to do so. In order to get back on track, we plan on working extra hours outside of class to get the model implemented by the end of the week.
By next week, I would like to have the initial model implemented so we can begin our first round of training and experimentation. Once we can establish an initial baseline for the model performance, we can start integrating biasing, hyper-parameter tuning, and specialized training. This iterative process should also take around a week and a half and will also include getting our cloud services up and running.
TEAM STATUS:
- What are the most significant risks that could jeopardize the success of the project? How are these risks being managed? What contingency plans are ready?
Right now, one of the biggest risks is simply staying on schedule. We have run into issues of not having parts ready because they needed to be ordered. Now that we have realized that this could potentially affect our timeline, we are reviewing any and all parts that might need to be purchased and making sure we have all of them by the end of this coming week.
- Were any changes made to the existing design of the system (requirements, block diagram, system spec, etc)? Why was this change necessary, what costs does the change incur, and how will these costs be mitigated going forward?
We have pivoted back to Faster-RCNN as we believe that the accuracy tradeoff from YOLO was not worth it’s inference speed. This does not affect out project too much although we did take a bit too long to decide between the models and now we are slightly behind schedule.
C9