This week we worked together to prepare for in class demos by polishing up the robot structure and creating a working prototype of our software system. Meghana and I worked on reducing the weight and friction of the robot by adding caster wheels beneath the structure, which allows the robot to move at essentially the same speed as it did without the entire structure on top. Mae and I worked on a pipeline for our software system by embedding the distance detection and navigation control within each inference run. Now, everytime the robot samples a frame, it’ll run inference, obtain bounding boxes for the bottle and obstacles in the frame, and if a bottle is detected, our algorithm will choose the closest bottle as the selected bounding box, and determine which side the bottle is on relative to the robot. It will then begin the navigation control which consists of sending a signal to the iRobot to turn in the direction to face the bottle head on until the center of the bounding box coordinates has reached a range of pixels in the vertical center of the frame. One frustrating aspect is that inference still takes 20+ seconds since we have not been able to get CUDA to work on the Jetson. We have attempted to flash a Jetson Nano but realized that Jetpack is not supported on the Nano and that we would be unable to access CUDA in that way. Our option now is to reflash the Xavier to access the JetPack update. Another software endeavor we worked on was to incorporate multi-object tracking still using CSRT so that we can track both bottles and obstacles and pause the motion of the robot if it sees an obstacle that becomes closer than the bottle being tracked.
Next week, I plan to work with Mae to continue completing the software pipeline to incorporate multi-object tracking so that we can start doing full system testing. We are still a little behind schedule, but making good progress with robot construction and the overall software pipeline.
0 Comments