Autonomous cars refer to the term of a vehicle that does not require a human to function yet is able to maneuver itself around streets and highways. They are the future of transportation and will change driving on roads forever.
There are six levels of autonomy in cars. Level zero refers to a car that has no autonomy and the driver has full control of the vehicle. This level would refer to older cars mainly from the 1990s. Level one refers to a car that has basic assistance such as ABS (auto brake system), ESP (electronic stability program), Cruise control, and other helpful but not fully autonomous assists. This level would refer to modern cars that most people drive on the road today.
Level two refers to a car that has some autonomy such as the car accelerating, steering, and braking on its own, but it has to be monitored by the driver. This level would refer to cars like those from Tesla.
Level three refers to a car that is autonomous enough so that the driver doesn’t even have to keep their eyes on the road in certain situations. Currently, there are no cars on the road that have these capabilities but many are being prototyped.
Level four refers to a car that drives itself in almost all situations without the driver’s help. This is the goal car that all autonomous car manufacturers are aiming for but not many have reached so yet.
Finally, level five refers to a car that drives fully autonomous and there is no need for a driver. This is considered the perfect autonomous car but it is understood in the community of autonomous car manufacturers that this will not be a reality anytime soon so the goal is set to level four autonomy.
There are eight main factors that make up autonomy. Mapping, route planning, perception, localization, prediction, behavior planning, trajectory generation, and control. These 8 factors all work together to make the vehicles autonomous and efficient.
Mapping, route planning, and localization are the main parts of the autonomy that allows the vehicle to understand where it is and where it needs to go. There are three types of maps that the vehicle utilizes when mapping which includes semantic maps, lidar maps, and radar maps. Semantic maps are AV-specific high-definition maps, with centimeter-level resolution. They have details on the exact location of intersections, crosswalks, and parking structures to allow the vehicle to determine whether a vehicle is parked and can be passed. These maps also provide the vehicle with the precise location of traffic lights and stop signs for each lane as well as speed limits, turn restrictions, and lane directions. The lidar maps are also generated by engineers to bounce maps of the entire drivable region. These lidar maps are used to determine the vehicle’s precise location. The map associates each latitude or longitude with the laser-ranged distances to major landmarks, such as trees or buildings. Radar maps use waves to detect objects by bouncing those waves out of the car and off of other objects back into the car to determine where certain objects are around the car and whether it is safe to continue moving or not. Route planning is a special part of the car’s autonomy for it is not a part of the actual vehicle. There are data centers that compute an optimal route for the vehicle similar to that on google maps when you are determining the fastest route to your destination. Routing instructions include details on which lanes the vehicle needs to be in, where the turns are and other common instructions of the type. The route can take into consideration real-time traffic conditions, construction, and other factors of the kind. Localization is a very important part of the location services of the vehicle. Localization is the term used to describe the vehicle determining where it is at the exact moment. The GPS provides roughly 5 meters of accuracy so it is not enough to determine the exact location. The vehicle is able to compare its own lidar readings to its pre-built lidar maps to figure out its location. IMUs (Inertial Measurement Units) measure the vehicle’s rotation, acceleration, and other specific data points providing extra accuracy.
Prediction, behavior planning, and trajectory generation, as by their titles, are some of the more important aspects of autonomy as they determine future events allowing the vehicle to react better and make the trip as safe as possible. Prediction is the system of the vehicle that predicts other vehicles' future behavior based on recent history. The system uses eclipses that grow progressively to show the uncertainty as they progress but is enough to predict the future trajectories. The autonomous vehicle leverages approaches like neural networks and probabilistic reasoning to generate these predictions for nearby agents, such as pedestrians, cyclists, and other vehicles. Behavior planning includes traffic elements, such as stops signs and speed limits, and static objects, like parked cars and boundaries. To do so it considers future predictions of dynamic obstacles. The behavior planner proposes one or more high level maneuvers for the vehicle which it then outputs as a set of space-time constraints that the vehicle must honor. Trajectory generation uses the given behavior plan to generate a preferred “kinematically feasible” trajectory for the vehicle. To do so, some trajectory generators generate 1000s of candidate trajectories, many of which include options that are infeasible, and then filter them against the space-time constraints generated by the behavior planner.
Perception and control are two major components of autonomy that allow the car to function at the exact moment such as steering, accelerating and reading certain road signs around it. There are three types of perception, camera, radar, and lidar. The camera perception uses cameras on the car to visualize objects and obstacles around itself. The radar perception uses waves similar to that of the radar mapping to find the location of obstacles but doesn’t give a visualization of the object. Lidar perception uses technology like that of lidar mapping to identify objects around it. Finally, the control uses all the given factors summed up in the generated trajectory to execute the car’s actions. This translates the trajectory into a set of steering, acceleration, and braking commands of the car. An algorithm called a PID controller is often used for this. Overall the controller is the final step that actually executes what the other sensors and data points have determined as the most feasible path.
Some other critical supporting roles are the simulations of certain situations, conversion of metrics, cyber security, validation to determine what is feasible versus infeasible, and machine learning infra, which allows the vehicle to learn from everything it does and calculates. The trolley problem is sometimes a topic that arises when discussing autonomous cars. This is a series of thought experiments involving ethical dilemmas of whether to sacrifice one person to save a larger number. The ethical area of study was dubbed the “trolley problem” in 1976 by Judith Jarvis Thompson. Although an autonomous vehicle, it will be able to hit the brakes as an alternative compared to trolleys who couldn’t as easily it still is a thought-about topic by many during these discussions. Overall, these are the main topics of an autonomous car and how they function.
—————————-
By: Zubin Sidhu
—————————-
References:
Lutkevich, Ben. “Self-Driving Car (Autonomous Car or Driverless Car).” SearchEnterpriseAI, 30 Oct. 2019, searchenterpriseai.techtarget.com/definition/driverless-car.
Comments