Building Perception, Planning and Control for Autonomous Vehicles
I chose to pursue my Bachelor Thesis project at IIIT Hyderabad, where I was given the opportunity to work on Mahindra e2o vehicle. The task essentially required various modules to be built. It can be divided majorly into the following areas:
Control
Perception
Localisation
Planning
Initially, myself along with other interns were put on the task of designing a low level controller to the vehicle. We tapped in wires from various parts of the vehicle like steering, throttle and used a low cost Arduino Mega to design control algorithms for them. In this process we designed a novel longitudinal speed controller which culminated into a paper "Gradient Aware - Shrinking Domain based Control Design for Reactive Planning Frameworks used in Autonomous Vehicles". Localisation was achieved by fusing various sensor information such as GPS, IMU and odometry from LOAM velodyne using lidar scans. Occupancy grid was created using the lidar pointcloud scans. The perception of the car was limited to this extent of knowing whether certain area in the scene is obstacle or not. For a start we started by implementing RRTstar based planner on the occupancy grid created using lidar scans. With this we exhibited a driver-less static obstacle avoidance in tight spaces. We are further extending to exhibit dynamic obstacle avoidance using stereo cameras by doing Time scale collision cone.
For further details, please refer to my Project report