Navigation for agriculture robots [TerraSentia robot, PyTorch]
Mentors: Girish Chowdhary and Anwesa Choudhuri
Agriculture is a very cool application of robotics. The robot will have to complete tasks in a GPS denied environment where everything looks the same, is slightly dynamic, with constant camera occlusion. When I first joined DASlab I assisted with two projects for the TerraSentia robot. The first project involved training a deep learning model to predict whether the robot is in a crop row or not. This was useful for other navigation and planning tasks performed by the robot such turning into a new row of crops. The model I trained ended up getting around 96% accuracy on this task.
The second project consisted of writing some scripts to do nightly testing of the robot’s software. This interfaced with some external hardware as well as a Gazebo simulation.