search

JacSchn / 5G-Beamforming-from-Visual-and-Lidar-Rendering

Regent Scholarship Research with UWW where we are using machine learning in an autonomous driving lab to collect visual and lidar-based models to have optimal guidance of 60GHz Wireless Network.
GNU Lesser General Public License v2.1
2 stars 1 forks source link

Computational Limitations of the Nano #23

Closed flynn248 closed 2 years ago

flynn248 commented 2 years ago

The Jetson Nano, although powerful, is not able to handle data collection and self driving all on one device. This became apparent when it was attempted and the vehicle failed to stay on the track even though it successfully stayed on when it was just doing the self driving. $ tegrestats shows the status of the memory, CPU, GPU, etc. of the Nano. Using this and incrementally turning on different sensors shows that the two USB cameras use up most of the CPU.

Luckily, we have two identical Nanos that were updated to the same spot in development. So, one Nano is dedicated to self driving while the other is dedicated to data collection. The camera for self driving is not one of the sensors that data will be collected from. So, it is not needed to run donkeycar in a ROS package. This does render the csi_drive_data package somewhat useless currently, but it may find a use in the future.

flynn248 commented 2 years ago

This issue was resolved by placing the second Nano above the first. The bottom Nano drives the vehicle while the top Nano handles all of the data collection.