Barriers to Achieving Fully Autonomous Vehicles
Where are all the autonomous cars?
This article was co-written and reviewed by Professor Fikes of Stanford University, as part of an internship during my sophomore year in high school. Thanks to his help, I was able to gain valuable insight into the world of autonomous vehicles, as well as speak to those working in the industry.
Introduction
Five Levels of Autonomy
When ranking autonomous vehicles (AVs), experts use a five-level system that ranges from AVs that assist human drivers with some driving tasks, to ones that drive autonomously in all situations. The descriptions of each level as defined by the NHTSA¹ are detailed below.
Level 1 Automation: Driver Assistance
AV can assist human driver with one driving task in some situations
Level 1 vehicles can perform one driving task in some situations. In most modern vehicles there exists some kind of Level 1 autonomous feature, most commonly being adaptive cruise control that monitors the speed of a vehicle in order to maintain a consistent speed and a safe distance from traffic ahead. The human driver needs to monitor the driving task the Level 1 vehicle is performing and be prepared to intervene when necessary. The human is still driving the vehicle but is assisted in minor ways.
Examples: Toyota, BMW, Mercedes, etc.
Level 2 Automation: Partial Automation
AV can assist human driver with multiple driving tasks in some situations
Level 2 automated vehicles have an automated system that controls multiple driving tasks such as steering and acceleration, but only in some situations while performing specific driving maneuvers such as parking. The human driver needs to monitor the driving tasks the Level 2 vehicle is performing and be prepared to intervene when necessary.
Examples: Tesla FSD², Mercedes-Benz Intelligent Drive
Level 3: Conditional Automation
AV can perform all driving tasks but requires human intervention in some situations
Level 3 vehicles are able to perform all driving tasks on the road under some circumstances. The most common example of this is automated highway driving. The vehicles still require human intervention if necessary, but do not require drivers to be as attentive as they would in level 2 vehicles. These vehicles exist in a grey area, where the autonomous vehicle is equipped to handle most situations, but still requires a human driver to stay attentive and take control when a low-frequency situation occurs that the vehicle cannot handle. Marketing such a vehicle as “autonomous” may cause drivers to be inattentive and lead to possible accidents due to mismarketing the capabilities of the product. For this reason, most AV companies have ruled out the possibility of commercially viable Level 3 vehicles.
Examples: Audi A8, Lyft, Uber
Level 4: High Automation
AV can drive without human intervention in prespecified situations
Level 4 vehicles are considered as fully autonomous, using an automated system to handle all aspects of driving, but only within prespecified situations. These limitations most commonly include things like geofencing, where vehicles are only able to operate within a certain region. For example, industry leader Waymo is only able to test its Level 4 vehicles within portions of Arizona suburbs³ due to these limitations. A level 4 vehicle need not require a driver to be aboard the vehicle, but may have driver controls (e.g., steering wheel and brake pedal) so that a driver can control the vehicle. Although Level 4 vehicles are restricted, companies are working to develop vehicles that can thrive in these limitations, like trucks with consistent, well-optimized routes, and taxis that only function during certain time periods and in restricted geographical regions.
Examples: Waymo⁴
Level 5: Fully Autonomous
AV can drive without human intervention in all situations
A fully autonomous vehicle would manage all aspects of driving by itself without any interference from a human driver. Unlike Level 4 vehicles, the vehicle would not be geographically restrained. Prototypes for these types of vehicles include designs that lack steering wheels and brake controls because a vehicle such as a self-driving taxi would not need any controls.
Examples: N/A
The issue currently plaguing the AV scene is the transition from level 4 vehicles to level 5. Companies have become more cautious in announcing new automated vehicles after the first fatal AV accident back in 2018.⁵ Timelines for companies releasing fully autonomous vehicles have been pushed back decades, caused by limitations in AI, sensory devices, and other aspects.
Currently there is no mass production of any vehicle higher than level 2. This paper will detail the problems surrounding the leap to level 5 autonomy, as well as various solutions to overcoming them.
The Driving Tasks
Driving is a continuous task of determining and performing the vehicle’s next actions that will move it toward its destination and will avoid dangerous events.
In order to do that task, the driver must (1) understand the environment in which the vehicle is located, and (2) predict the movement of entities in that environment that are relevant to determining the next actions of the vehicle.
Human drivers understand environments through their senses, most notably vision and hearing, as well as knowledge of the location. This includes recognizing places like crosswalks, then applying their knowledge from previous experiences to drive.
An AV senses its surroundings using sensory devices that most commonly consist of cameras, radar, lidar and GPS.⁶
This sensory data is useless to the AI driver unless it is able to be interpreted and understood, just like how humans use their brain to make sense of what they see. Technology like image recognition software is necessary to classify objects like pedestrians or vehicles, as well as record their approximate location. Through this process, the self-driving AI is able to create what is commonly referred to as “computer vision” for the AV.
Simply knowing the locations of the objects around the vehicle is not enough to ensure a safe drive. Car accidents occur due to the collision of two entities, so human drivers subconsciously infer where cars are likely to be headed, and plan their route accordingly.
Humans are remarkably good at this task, but AI has much more trouble accurately predicting the routes of other vehicles.
Next, the AV needs to create a route to its destinations where it is able to identify and avoid any possible collisions. This identification and avoidance comes in the last step of the self-driving process, motion prediction. This step includes using the computer vision created previously and judging where vehicles and other objects will move based on their previous trajectories, speed, and behavior.⁷ Once a motion prediction model is made, the AV can pick which possible route is the safest and does not collide with the predicted trajectories.
What Can and Cannot Be Done Automatically
Understanding the Environment Around the Vehicle
In order for the AV to create a safe course of action to arrive at the passenger’s destination, it must first collect and understand information from its surroundings. This process is achieved through physical sensors placed on the AV, data exchanged from the cloud or other neighboring vehicles, and several AI programs.
The AV’s Physical Sensors
In order for a vehicle to reach its destination, it needs to first sense its surroundings. This is done currently by cameras, radar, lidar, ultrasonic sensors, and GNSS.
GNSS, more commonly known as GPS, is used for determining a vehicle’s location using satellite technology. Along with navigation systems like Google Maps, an AV can use its location to determine the most efficient route to a destination. GPS accuracy can be deteriorated by factors such as weather but has remained as a reliable navigation tool for decades. Under ideal conditions, where there are not large structures such as mountains or towers where the GPS signals may be interrupted, GPS can reach an accuracy up to 2.5 meters, while a weak signal may only reach an accuracy of 10–15 or more meters.⁸ Luckily, the GPS itself is able to tell if it’s signal is weak or being obscured, and a vehicle can still estimate it’s approximate position using maps without relying on continuous GPS data.
Ultrasonic Sensors are time-of flight sensors that utilize inaudible sound waves to detect obstacles. They have a limited range and are most commonly used to assist with parking to alert the driver when the vehicle is close to an object. These sensors can guarantee a detection of an object within 10 feet, but unreliably detect objects within 15 feet.⁹ They are typically placed on the front and back of a vehicle to warn the AI or human driver of possible collisions.
Cameras, or more specifically 360° optic cameras, serve as the “eyes” of the vehicle and are used to detect and recognize objects.¹⁰ Other sensors like radar are able to reliably detect objects, but cameras are best due to their ability to provide visual data that can be processed by AI classification systems to distinguish objects like vehicles, pedestrians, street signs and crossings. However, in situations that any human driver would not be able to see well, such as rain or fog, optic cameras will have trouble seeing as well. Some AV’s have infrared cameras that help with obstacles like darkness, but weather conditions still remain a problem. Without objects obscuring line of sight, companies have reported an accurate detection of traffic signs in a maximum range of 200 meters.¹¹ The largest limitation of the cameras is their inability to reliably map depth. Cameras deliver a flat image which can be processed through depth-estimation software constructing a 3D representation of the AV’s surroundings. The current software still results in a poor reconstruction with missing or noisy borders and is unable to perform well when faced with reflective or dark surfaces.
Radar is also used frequently in automated vehicles for its ability to detect metal objects.¹² This is helpful for detecting objects like cars, but unlike cameras, radar is unable to consistently distinguish cars from other unexpected metal objects. Companies have cited obstacles like broken guard railings and street signs as issues for the radar sensors that would misinterpret these objects as vehicles. Most AV’s are equipped with multiple radar devices, usually long-range radars placed on the front and rear that can detect objects up to 200 meters ahead, and short-range radar on the sides of the vehicle that detect objects roughly up to 20 meters from the vehicle.
Lidar has been claimed to be essential by most Level 4 developers, but it has been dismissed by other companies developing level 2 vehicles due to its high price point.¹³ There is agreement, however, that Lidar is essential for AV’s of higher levels. Lidar can be used to create accurate mappings of its environment by using time of flight measurements to find the distance between the objects around it. To tackle the high price point issue, solid-state lidar sensors have been introduced that are much cheaper because they do not include mechanically moving parts. Lidar shares a lot of the same limitations as optic cameras and its line of sight can be easily obscured. Lidar can detect objects as far as 200 meters from the vehicle.
Interpreting the AV’s Sensor Data
The usage of artificial intelligence in AV sensor systems has been crucial for recognition.¹⁴ AI camera image recognition technology has already been used for facial recognition, disease detection, and various other uses. By using AI, a system can be trained from recorded sensor data to recognize and classify objects as cars, trucks, and pedestrians. This makes camera data especially useful, as the AI system may also attach predicted behaviors to things like large commercial trucks, which may move slowly and allow smaller vehicles to pass them. Lidar can also be used to classify objects, but it is not nearly as powerful as camera recognition.
No sensor is 100% accurate, and they each have their own limitations. Because of this, all AVs use the strategy of sensor fusion that combines the data collected from all of its sensors to “cross-check” the existence and location of an object.¹⁵ This is extremely important because it provides a failsafe for each of the sensors in case it is presented with an environment where it can no longer function properly.
The various ways that AV sensors can fail will cause inaccuracy in the data overall. Luckily, multisensor data fusion is not a problem exclusive to the development of autonomous vehicles, and techniques exist for effectively dealing with this issue.¹⁶ AI sensor systems can recognize repeated inconsistencies in sensor data and account for uncertainty in each sensor. This means if a sensor like a camera was facing harsh weather conditions, the AI system could recognize it’s inconsistent data and choose to rely more heavily on a sensor like Lidar when constructing a model of its surroundings.
Data From other Sources
One of the ways that AV companies are increasing the accuracy of computer vision is the development of high definition (HD) maps of driving environments.¹⁷ By combining data collected by AVs from sets of cameras, LiDAR, GPS, and other sensors, a detailed model of a fixed portion of an environment can be created and made available to each AV that enters that environment. This could be used to significantly aid the data the AV can work with. When there is uncertainty in the data from an AV’s sensors due, for example, to extreme weather conditions, the AV could rely on these HD maps to identify objects that could not be recognized by their degraded sensors. These HD maps are what allow companies like Waymo to succeed in a limited geographical area where they have created highly detailed HD maps. The problem with these maps is that they are expensive to make and store. The maps are based on the data collected from AVs, so any flaw or misinterpreted object would also be added to the HD map. That is why AV makers have had to rely on human verification to fix the errors in the maps, which is costly and time intensive. Another issue with relying on HD maps is having to stream the data to the AV continuously. The maps store an estimate of one terabyte of data a day, a lot of sensor data, and without an improvement to network speeds like 5G, the current transmission speeds are not fast enough for the AV to fully interpret the data and utilize it. It could be a possibility that AV’s could download the maps before the commute is made, but this download must be planned ahead of time. Unexpected rerouting and changes of destination still have to be taken into account, meaning that an efficient way to store and transmit HD maps still needs to be implemented.
Other maps besides HD environment maps are also being used to aid AVs. Traffic maps can allow the AV to predict the general flow of traffic in a given area, which would aid the AV in the motion prediction process. There are also maps that allow the AV to identify parts of the world such as intersections, parking lots, highways etc., and then use prior knowledge to predict what types of situations could occur in an area.
Vehicle-to-vehicle (V2V) communication,¹⁸ found in various modern vehicles, allows vehicles to inform other nearby vehicles of their speed, direction, and location, making less of their predicted movement variable to the AV’s AI system. Current V2V communication is limited to a range of 200 meters. We are still a long way from having all vehicles on the streets containing V2V technology, and, even then, pedestrian prediction will persist as a major issue.
When driving in new weather, lighting, and environmental conditions, it is important to factor in the changes in camera sensor data that occur. The AV’s computer vision is extremely dependent on training the AI system on data, so driving an AV in fog or a different locale that it has not experienced could cause various issues. Due to this issue, experts have used image-to-image Generative Adversarial Networks (GANs) to translate images collected from a familiar environment into the style of the new environment.¹⁹ This allows the vehicle to apply its prior knowledge from old locations to understand and adapt to new ones by recycling the data from old locations in order to successfully interpret their surroundings in new locations.
Understanding What’s Going To Happen Next (Motion Prediction)
Predictive Models
When creating autonomous vehicles, one may think that creating a route for the AV to follow would be very simple. If given enough sensor data to understand it’s current surroundings, the car must be able to follow the same rules taught to new drivers at the DMV in order to drive safely.
As it turns out, a rule-based approach like this would not work in a real-life driving setting. There is a countless amount of etiquette, common sense, and real-time decision-making skills that must be acquired and that are not stated in any driving rule book. This is the same reason why humans are required to practice driving with an adult before getting their license. Rule-based AVs would undoubtedly thrive under predictable conditions like parking lots and highways, but when faced with situations like jay-walking and other situations that are out of the ordinary, the AV may respond poorly. When relying only on hard-coded human created rules, it may be unable to respond properly when it sees other vehicles act out of this status quo or in situations that don’t have the same predictable behavior.
The most common way autonomous vehicles are being tested is through simulations. These simulations are able to create an environment with vehicles, streets, and pedestrians in it for the AV to respond to. It allows AV developers to track their performance through running a simulation and thereby reduce costs of building vehicles and eliminate the possibility of any real-world harm. However, creating safe autonomous vehicles needs more than just thousands of simulations to be run and validated so the AV can be updated. Firstly, it is nearly impossible and inefficient to simulate infinite amounts of situations. No matter how many simulations are run, there will always be unpredictable ways that vehicles and pedestrians will behave and interact with one another. Secondly, the AV still needs some kind of reasoning skills beyond what is provided from any human validator and be able to apply these skills in real time.
A standard approach to this problem, which companies like Waymo and Lyft have taken, involves taking a large dataset of sensor input and using it to train an AI system to create motion predictions using deep learning.²⁰ This means, unlike simply relying on the snapshots of information provided by the sensors, the AV will be able to create a predictive model representing the least and most probable movements neighboring pedestrians, vehicles, and other objects may take in the future. This approach has been successful and widely adapted, allowing Waymo to achieve Level 4 vehicles within parts of Arizona, although only in small selected suburbs.
Motion prediction is considered to be the most important part of the self-driving process. Instead of simply coding in strict rules for the computers to follow, predictive models allow the AI to have a form of reasoning skills, where they are able to be trained off of past data from driving in order to accurately predict how likely a car is to move in certain ways.
For example, if the AV’s sensors detect a truck to the right of it that can either turn left or continue forward, it will assign a percentage representing how likely it thinks each action could occur. Based on this, the AV would determine a course of action that would get it from point a to point b that avoids collisions from either possible action. If an action had a very low predicted percentage of occurring, it may be taken into less account than the most likely action. Real models are much more complicated than these, with thousands of different predicted trajectories for each vehicle being modeled, as well as other vehicles, pedestrians, and objects it must accurately predict and avoid collisions with.
Use of predictive models creates proactive driving,²¹ where AVs can change their models of motion prediction in real time, contrasted to reactive driving where AVs would only respond after a vehicle had made an action, allowing the AV to prepare for any possible trajectory that another vehicle could take. Since the AV is able to account for actions that could happen before they occur, these vehicles could respond fast enough to previously unexpected u-turns or jaywalking, which would be represented in the model.
The rules given in each area may still be present, but would only serve to aid the accuracy of the predictions. It is important for the AV to recognize structures like lines indicating a parking lot and be able to behave in ways that a car needs to in situations like these. However, another approach to this could be remembering and sharing behaviors found in places like these, which could lead to more accurate predictions that do not factor in biased hard-coded rules from developers. When only using rules created by developers to aid in AV situations, you are only able to factor in perspectives from a handful of people’s experiences, which is limited compared to an AI system that factors in millions of patterns it has picked up which a small group of people may not realize.
At this point in time, companies have reported producing AVs that are able to competently do motion prediction in over 99% of all situations.²² The industry is struggling to reach closer and closer to 100% safety in order to reach a point where such Level 5 vehicles could become commercially viable. Waymo believes that filling the gap in safety involves collecting more and more sensory data to create detailed HD maps and detailed lidar maps that can be reused and shared between vehicles. However, other approaches involving ties to cognitive science could be the key to safer autonomous vehicles. In the next section, we will be looking at how various companies have been identifying social aspects from drivers and pedestrians in order to improve the existing motion prediction models.
Social Prediction
In recent years, social prediction has been declared essential by more and more AV companies. Social prediction in AVs is the usage of social factors in order to improve the accuracy of motion prediction models, beyond simply analyzing the positions and trajectories of surrounding objects. For most companies this means determining factors like the driver of another car’s awareness of the AV, its intention, or it’s behavior.
MIT spinoff iSee has been taking a different approach to closing the gap between 99% and 100% by attempting to engineer a type of “common sense” into their AI systems.²³ By this, they mean that they are applying models from psychology, specifically the Bayesian Theory of Mind. The Theory of Mind is a skill that develops very early on in humans and is the ability for a human to recognize the intentions of other humans. When applied to AVs, they use vehicular motion to recognize the intention of drivers of other vehicles and pedestrians. For example, during a merge, a car needs to establish priority for the other to allow it to merge. None of this communication is done verbally, so the AV must be able to recognize patterns that signify that the driver of a car wants to merge.
Analyzing mental states of drivers and pedestrians would allow for a second layer of prediction above the mentioned prediction models and could be the key to making much more accurate predictive models. These mental state inferences in which we subconsciously assume intention and awareness of other drivers occur frequently during driving, and our proficiency in making them are one of the reasons why humans are such good drivers. AVs may be able to remember and share recognized vehicular motion, further strengthening their ability to predict the trajectories of other vehicles during complex social interactions.
These patterns which correspond to certain mental states of drivers are remembered in the AV’s AI system and can be shared with other AVs. This plays into why so many AVs today are geofenced, as it will take a substantial amount of time for each locale’s vehicular motions to be identified and updated. Understanding these more social elements of a driver or pedestrian allows one to more accurately predict motion prediction that is not based solely on speed, trajectory, and human-made rules, but associates vehicular and human motion to agent objectives and behavior. This strategy of finding why vehicles behave a certain way, rather than just finding more and more data, could be the key to creating Level 5 autonomy.
Since these mental state inferences are so essential to safe autonomous driving, it is also important for AVs to demonstrate their intentions back to human drivers. Maneuvers remembered by the AV’s AI system can be replicated so human drivers can, for example, infer that an AV is establishing priority by moving forward and decide that they should yield.
Researchers at MIT’s CSAIL have identified important classifications in the behavior of human drivers.²⁴ Most commonly, a driver can be identified as altruistic, egotistical, and prosocial. By identifying a car as egotistical through vehicular motion, they can determine that the car is most likely to repeat an egotistical action such as impatiently merging, and also how neighboring cars are most likely to react to it. They have reportedly reduced motion prediction errors by 25%, as well as provided a way that can better explain why an AV acted in a certain way.
These social patterns are not universal however, as drivers may display different types of vehicular motion to indicate something depending on the customs of their region. For example, some vehicles will act much more impatient than they would in different regions. And, in a country like China, an AV would not be able to utilize any hand signals given by the driver, as things like waving or eye contact are not as prominently used for communication as they are in the United States. Creating AVs that analyze mental states will require AVs to identify which patterns are universal, and record which are specific to distinct regions.
Pedestrian Social Prediction
In terms of accurate motion predictions, pedestrians are the most unpredictable. Cars can be expected to follow the status quo and behave a certain way while driving on the road, but humans can act in millions of ways that the vehicle will not be able to accurately predict and map.
Instead of only collecting data from pedestrians like speed and direction, research suggests that the same reversed theory of mind model needs to be applied to pedestrian movement prediction. Instead of vehicular motion which is identified in vehicle prediction, factors like gesture (body language) and eye contact are used to create inferences for pedestrian prediction Based on these action inputs, the AV can determine two important factors about a pedestrian — their awareness of the AV, and their intention to take some kind of action like crossing a road. A company known as Perceptive Automata²⁵ specializes in these pedestrian systems, partnering with investors like Uber and Toyota. Their models can detect things like a pedestrian on their phone, and based on that inference they can make maneuvers which don’t require the attentiveness of the pedestrian. On the other hand, their models can recognize a pedestrian checking left and right on a crosswalk attentively indicating their intention to cross and prepare to stop for them.
2.3 Deciding What To Do Next
What separates level 4 vehicles from level 3 vehicles, is their ability to function without a human failsafe. When level 4 vehicles encounter a situation that they are unable to handle, they are able to put themselves in a safe position and then allow the human driver to take over. This means that a fully autonomous vehicle needs to know what to do in every driving scenario it can be put in, including the case of an unavoidable accident.
There is an interesting moral issue that comes along with developing AVs. Since there is no such thing as a 100% safe AV, the AI system should be capable of doing “damage control” when certain that it will cause an accident. The issue is similar to the trolley problem, which asks a participant if they would witness the killing of 3 people or interfere and cause the death of 1 person.²⁶ In a fatal scenario, there could be issues with bias programmed into the AI system causing factors like age, race, and occupation to be factored into who would be chosen to live. It becomes the industry’s responsibility to ensure no bias is programmed into their machines, or controversy could arise.²⁷
Although this utilitarian approach seems like the obvious choice for dealing with fatal situations, protecting the passenger above all else needs to be considered. Autonomous vehicle customers would more likely purchase the car that protects themselves in the case of an inevitable accident, rather than sacrifice itself to save others. Their actions also have to be explainable. An AV might do an unusual maneuver that the AI system predicts would have the highest likelihood of avoiding an accident. In the case that the accident is still not avoided, the developers of the AI system need a way of explaining why it behaved in that way, which may be difficult if we can’t decipher the vehicle’s decision-making process.
Another important factor that must be considered when making AVs is that customers may pick an AI system that aligns with their own beliefs. For example, the majority of customers would pick the vehicle that is not as utilitarian, and when predicting an oncoming collision will always favor the passenger inside the vehicle. Another more selfless customer may choose something that always picks the outcome which avoids the most number of casualties. AV developers may be incentivized to only create vehicles that protect the people inside it, as it can be marketed as the most safe.
Conclusions
Major strides have been made in the last couple of years towards achieving Level 5 autonomy, but we are still possibly decades away from the release of any mass produced commercially viable products with that capability. The largest roadblock comes in motion prediction, where there are near infinite amounts of possible situations, with companies either looking for more ways to obtain accurate sensor data or creating new ways to teach the AI system.
We can draw a couple of conclusions from the topics discussed earlier:
In terms of accurate motion predictions, pedestrians are the most unpredictable.
The usage of artificial intelligence in AV sensor systems has been crucial.
AV’s need to have more than just rules, they need learned neural network prediction models to become autonomous.
Social Prediction may be necessary in creating accurate motion prediction.
Things like social prediction have provided “another layer” of accuracy on top of motion prediction and rule based systems, but it is still up to debate if this is enough to bring autonomous vehicles to a safe position. The percentage of safety that is acceptable may vary from region to region. Countries like China have been early adopters of potentially risky technologies, but countries like the U.S. may be more wary, and customer acceptance may only accept an AV which surpasses the capabilities of the best human driver. In terms of infrastructure and policy, the U.S. is quite ahead as it has the most major AV companies, it has already been mapped visually by services like Google Maps, and it’s government has already allowed testing of level 4 vehicles.
What Can Technology do now?
As stated in earlier, Level 2 and Level 4 vehicles are still extremely viable and have made promising releases recently. Level 2 vehicles do not need to follow the restrictions of the higher level vehicles since they have a human driver who is responsible for the vehicle’s behavior, but they can be in full control for certain tasks like automatic parking, and in locales like highways and suburban streets. Level 2’s can be treated as non-autonomous vehicles with features of an AV. They can have features like valet parking and possibly self-steering features, but don’t have to be as functional as a real level 4 autonomous vehicle as they are always able to rely on their attentive human driver when faced with uncertainty.
For Level 2 vehicles, there are many AVs out there that are able to assist or take over various aspects of driving, which we can predict to become more and more common in the upcoming years. Level 2 encompasses a large variety of vehicles, as some features are as simple as adaptive cruise control and self-parking, while some are as advanced as Tesla’s FSD beta.²⁸ Tesla’s FSD beta can take over with full self-driving in situations it deems viable. Tesla stated it would allow its self-driving beta to be available to the public late 2020 for a subscription fee, a date which has been pushed back further.
Just a few years ago, Waymo launched their self-driving Level 4 Waymo One²⁹ taxis that operate exclusively in specific districts in Arizona. The vehicles allow passengers to pick their destination within a carefully selected 50 square mile area, and the vehicle will handle all aspects of driving without a driver in the front seat. However, their status as a level 4 AV can be put into question, as they have reported that uncertainty in situations is reported to humans who watch over the vehicles and who must respond quickly to ensure the safety of the vehicle. Having a group of human verifiers may work with a small fleet of taxis, but it is clear that this issue of uncertainty needs to be fixed when AVs are produced in scale.
Other applications include campus vehicles and self-driving trucks, which can all cut costs drastically in the transportation industry. These practical lower-level uses are able to bring us much closer to achieving Level 5 vehicles; Waymo’s taxis reported a major decrease in autonomous vehicle interventions just in a few years. Although Level 5 vehicles aren’t ready just yet, we can expect lower-level vehicles to incorporate self-driving into our daily lives.
Self-driving trucks³⁰ work especially well, as the majority of heavy cargo trucks’ jobs are following a straight path along long roads with very little interaction with pedestrians and other vehicles. These straight paths can be extensively mapped out, and the predictable behaviors of cars in highways will allow AV’s to function before the perfected models of social prediction . Truck driving is the most common job in 29 states, meaning that self-driving vehicles could skyrocket in revenue if these jobs were automated as automating these drivers could cut major costs.
The mass production of Level 5 vehicles will undoubtedly revolutionize our lives. The change to AVs will not only reduce traffic, and the average time of a commute, but also reduce the number of car accidents that occur. In 2018, 29% of all car crash fatalities were caused by drunk or distracted drivers,³¹ a statistic that could be harshly reduced with the rise of
AVs. Every aspect of transportation vehicles can be optimized and the costs of any previously required human driver would no longer be required. Those with disabilities or other factors that prevent them from driving can have a chance at becoming more independent. In a few decades, we may think of driving as something not only done by adults, as a driver’s license may not even be required depending on legislation. In the end, public acceptance will decide how long it takes for these AVs to be incorporated into the public. Given how AI is being depicted in pop culture and tech companies’ misusage of data being brought to light, it will take a long time for AVs to be ready to be released. Consumers may not realize the safety benefits from adopting an AV with a few flaws, and instead will hold AVs to unreasonable standards.
Sources:[1]: https://www.nhtsa.gov/technology-innovation/automated-vehicles-safety
[2]: https://www.tesla.com/support/full-self-driving-computer
[3]: https://waymo.com/waymo-one/
[4]: https://spectrum.ieee.org/cars-that-think/transportation/self-driving/full-autonomy-waymo-driver
[5]: https://www.bbc.com/news/world-us-canada-43497364
[6]: https://towardsdatascience.com/your-guide-to-ai-for-self-driving-cars-in-2020-218289719619
[7]: https://arxiv.org/abs/2011.03635
[8]: https://www.gps.gov/systems/gps/performance/accuracy/#:~:text=The%20government%20provides%20the%20GPS,interval%2C%20with%2095%25%20probability.[9]: https://medium.com/@BabakShah/ultrasonic-sensors-in-self-driving-cars-d28b63be676f[10]: https://becominghuman.ai/computer-vision-applications-in-self-driving-cars-610561e14118
[11]:
https://www.tesla.com/autopilot#:~:text=Advanced%20Sensor%20Coverage,to%20250%20meters%20of%20range
[12]: https://navtechradar.com/explore/autonomous-vehicles/
[13]: https://towardsdatascience.com/why-tesla-wont-use-lidar-57c325ae2ed5
[14]: https://towardsdatascience.com/how-do-self-driving-cars-see-13054aee2503[15]: https://mobile.aau.at/~welmenre/papers/elmenreich_Dissertation_sensorFusionInTimeTriggeredSystems.pdf[16]: https://news.mit.edu/2020/neural-network-uncertainty-1120[17]: https://medium.com/waymo/building-maps-for-a-self-driving-car-723b4d9cd3f4
[18]: https://www.nhtsa.gov/technology-innovation/vehicle-vehicle-communication
[19]: https://medium.com/@hello_isee/addressing-the-long-tail-of-autonomous-vehicle-perception-f0232ba2c6d4[20]: https://medium.com/@hello_isee/learning-to-predict-other-vehicles-trajectories-for-safe-autonomous-driving-cbfb1b8053a4[21]: https://www.youtube.com/watch?v=I2zNyPuQ-Bw
[22]: https://optometry.berkeley.edu/are-we-ready-for-self-driving-cars/
[23]: https://www.isee.ai/
[24]: https://news.mit.edu/2019/predicting-driving-personalities-1118
[25]: https://www.perceptiveautomata.com/
[26]: https://theconversation.com/the-trolley-dilemma-would-you-kill-one-person-to-save-five-57111[27]: https://www.moralmachine.net/[28]: https://www.youtube.com/watch?v=r7hsUkc3l_4
[29]: https://www.theverge.com/2018/12/5/18126103/waymo-one-self-driving-taxi-service-ride-safety-alphabet-cost-app[30]: https://www.atbs.com/knowledge-hub/trucking-blog/self-driving-trucks-are-truck-drivers-out-of-a-jo
[31]: http://cyberlaw.stanford.edu/blog/2013/12/human-error-cause-vehicle-crashes
