Brian Njuguna

 SAFE DECISION MAKING FOR AUTONOMOUS VEHICLE WITH SAFETY VERIFICATION Name of Student Course Tutor University Location Date Abstract With the development of some technologies such as artificial intelligence, autonomous vehicles will be widely used in future smart transportation systems and play an important role. The decision-making of an autonomous vehicle is challenging because of the uncertainty of the driving intention of the surrounding vehicles, the noise error of the sensor, and the perception blind zone caused by occlusion. Reinforcement learning has great application prospects for solving intelligent decision-making. It learns driving strategies through trial and error learning. However, this kind of intelligent decision-making based on reinforcement learning lacks safety guarantees and may be dangerous. This study seeks to provide a safe state of art decision making for autonomous vehicle with safety verification which will be achieved by exploiting reinforcement learning and intelligent decision-making techniques with real-time security verification. Key Words: Autonomous driving; safe decision making; deep reinforcement learning; safety verification Introduction Automotive safety is an important factor in the newly emerging phase of intelligent transport system, which includes innovation and new opportunities. A fireball of changes could wreak havoc on industry dynamics, introducing new participants and changing the roles of old ones. Even so, the above potentially transformative technological advances may have to counteract some obstacles (Mirchevska et al.,2018). The long-standing challenge of vehicle safety is to eliminate or mitigate the severity of potential casualties. Passive and active efforts are the two main types of safety-related efforts today, with the former being the more mature concept and technology. The universal automotive industry is already pushing the envelope to create advanced technologies to make driving safe, dependable, and self-sufficient. Now that the focus is on road safety, the industry is working together to create a prospect of negligible collisions and occurrences, as part of the Vision Zero initiative. People's perceptions of safety technologies have shifted as a result of recent technological advancements. From speed alerts to developed braking capabilities, technological advancements have made vehicles much safer, smarter, and more dependable. Productive automated vehicles (AV) technologies have the potential to transform a wide range of industries, including automotive, transportation, energy, agriculture, and others. Advanced computer systems technology, which is at the center of these platforms, is being heavily invested in order to realize this potential (Li et al., 2017). As a result, self-driving capabilities, such as those provided by automated driving systems and other automation solutions, are becoming more widely available. As the use of autonomous vehicles becomes more common, concerns about their safe operation and viability arise. In this study, autonomous decision-making is defined as a process in which selection has the capacity to comprehend a problem or choose a goal, and then make decisions under their own voluntarily or predicated according to their own perceptions to solve the problem or achieve the goal. In general, there are two types of decision-making methods: classical and learning-based methods. Typically, autonomous vehicles operate in a complex, dynamic environment while collaborating with other road users. Because traditional methods aren't always effective in such a driving environment due to their lack of robustness, learning-based methods are being used to improve autonomous vehicle decision-making (Iberraken, Adouane & Denis, 2018). In addition, with the development of new cognitive computing technological innovations, learning-based approaches have experienced tremendous growth and progress in the area of autonomous vehicles. The goal of a judgment program is to produce human-like safe and high-reliability driving strategies (Leitner et al., 2020). To do so, a set of design criteria must be developed, and five of them are outlined below. Good real-time decision-making performance; stabilize among driving effectiveness and security (usually, safety takes precedence over system performance); rational and correct generated decisions; vehicle ride comfort (gearshift consistency, less emergency braking); high fault detection capability. While (deep) RL techniques in particular do not guarantee safety or stability, there has recently been interest in researching RL-based control systems that seek to maintain some sense of safety throughout training and/or inference, such as to appease capacity boundaries or prevent unnecessary negative outcomes (Mirchevska et al.,2018). Methods to ascertain "relatively secure" domains of the state technology which makes smoothness inferences on the aspects, techniques that incorporate constructs from RL and regulation hypotheses frameworks predicated on reliability assessment logics, and methods to link some (discounted) value function identifying breaches of standards and specifications are among them. While the definition of "safety" varies depending on the context, several of these previous works include some structure of variation perception inside the deep RL loop, which is relevant to the current work. Literature Review Since its extensive application prospects in both military and civil fields, autonomous vehicle technology has steadily become the center of study as science and economy have progressed. Super intelligent automated military trucks can certainly aid soldiers in a variety of tasks like intelligence acquirement, fire strikes, and supervising (Mirchevska et al.,2018), while automated driving owns a brilliant ability to lower road accidents and mitigating congestion problems in the civilian sector (Leitner et al., 2020). A self-driving car is a complex intelligent system that combines technology for environmental awareness, path planning, decision-making, and motion control (Li et al., 2017). Decision-making systems are responsible for the effective and effective operation of autonomous cars as the "brain," and how to create high intelligent as well as reliable decision systems is gradually becoming the focus of researches in the area of autonomous driving (Schwarting, Alonso-Mora & Rus, 2018). The decision-making process is expressed in terms of generating human-level safe and appropriate driving behaviors while taking into account the surrounding environment, the movement of other traffic participants, and the adaptive control of ego vehicles; the resulted driving behaviors are then factored into the equation mostly by gesture recognition system to achieve efficient autonomous vehicle operation. DRL, or deep reinforcement learning, is a promising method for learning navigation in difficult autonomous driving scenarios. However, when it comes to building automated systems work in human situations, detecting the small signs that can imply significantly different outcomes is still a work in progress. This paper illustrates how directly insinuating the underlying condition and incorporating spatial-temporal correlations inside a reinforcement learning model can help with this problem. Reinforcement learning (RL) is a technique in which a computer software learns from its mistakes in order to achieve competitive advantages on a given job (Selvaraj, Ahrendt & Fabian, 2019). Supervised learning, unsupervised learning, and reinforcement learning are three main categories in which machine learning algorithms are frequently classified (RL). Unsupervised learning encompasses methodologies such as clustering algorithm or grouping implemented to unlabeled data, whereas controlled learning approaches are based upon inductive reasoning where the system is often developed employing labelled data to achieve classification or regression. In the RL paradigm, on the other hand, an autonomous agent interacts with its environment to enhance its effectiveness at a given task. In the areas of operating policy, predicting perceptions, route as well as motion control, and low-level controller design, RL is a potential option. The perception section's purpose is to create an intermediate level abstraction of the system's state (for example, a bird's-eye view map of all barriers and agents) that will be used by a decision-making system to generate the driving policy. This state would comprise lane positioning, drivable zone, agent location (cars and pedestrians), traffic light status, and other factors. Perceptional inconsistencies spread across the information chain. Because robust sensing is essential for safety, using several sources boosts detection confidence. This is accomplished by combining numerous sensory tasks into a multi-task model, such as segmentation, motion detection (Al-Nuaimi et al., 2021), distance estimation (Iberraken, Adouane & Denis, 2018), soiling identification (Selvaraj, Ahrendt & Fabian, 2019), and so on. Autonomous vehicles must navigate their surroundings while only having a limited view of other things on the road. Intelligent control errors, restricted sensor coverage because of weather or object detection lag, occlusion, and hidden characteristics as well as other human driver intentions are all sources of variation in autonomous vehicle measurements. When deciding on future vehicle moves, behavior planning must take into account all sources of uncertainty. Another key feature in this vehicle is the Driver alert control. It is a very unique technology which solves crisis that are occurring due to drivers who fall asleep during driving. The driver alert control system is developed to alert the driver in case they try to loose attention during driving on the main ways and highways (Selvaraj, Ahrendt & Fabian, 2019). It uses a camera that continuously monitors the movement between the road marking and asses if the vehicle is being driven in a right and controlled way. There is an indication on the dashboard to alert the driver in case of the reduced concentration and gives a message that recommends the driver to stop and take a break. Adaptive cruise control system utilizes radar sensors to monitor continuously the movement of the vehicle in front of a vehicle and adjusts the speed to the required limit to maintain the required distance (Iberraken, Adouane & Denis, 2018). Another method is the use of distance alert gives the driver information to make a safe distance since the system cannot accurately measure the exact distance between the car behind and the car in front. Related Work Both academia and business are still grappling with how to safely deploy autonomous vehicles in traffic. As a result, several industry activities and research on evaluating AVs exist. However, the majority of these approaches fall short in some way. The most popular type of AV evaluation is 'Shadow Driving,' in which a motorist is ready to intervene to avoid accidents or take control if the AV fails to engage (Li et al., 2017). The situations they concentrate on are predicated on both Berkeley's PATH-designed vehicle behavioral skills, including as identifying bikes and pedestrians and reporting to emergency services, along with Waymo-designed behavioral competencies. PEGASUS, a collaboration between different research and industry groups, intends to provide a whole instrument sequence for AV evaluation, drawing inspiration from both traditional vehicle verification as well as emerging breakthroughs in the sector. There have also been a number of recent projects aimed at standardizing AV validation by combining various methodologies. PEGASUS and Intelligent Testing Framework are two notable examples of such techniques (Mirchevska et al.,2018). There are additional simulators dedicated to vehicle dynamics, which allow one to evaluate the inner structure and physics of a car while it's on the road. V-REP is a vehicle dynamics robotics testing environment that allows users to play around with sensing devices, actuators, and optimization techniques (Leitner et al., 2020). A small number of simulators have been improved as profound learning technologies, focusing on the training procedure. NVIDIA's DRIVE Constellation uses two servers to quickly train the AV, one simulating realistic sensor data and the other simulating the whole AV software stack. By integrating True Life 3D Mesh and Physics studio, Cognata creates a full AV feedback mechanism besides a real context to teach AVs (Selvaraj, Ahrendt & Fabian, 2019). There are many works in the autonomous driving literature that look at risk-averse driving and risk assessment. Prediction is often used for safety in autonomous driving and accurate prediction models are a current topic of research in the autonomous driving community. The use of reinforcement decision making, artificial intelligence has proven to be the best solution to fill the gaps in automotive industry (Selvaraj, Ahrendt & Fabian, 2019). With illustrations and already existing examples as case study, the research will present a clear understanding on the safety decision making with data driven technologies. Driver alcohol detection system has been adopted by Volvo XC60 crowns have used the Alco guard in taking responsible decision to avoid driving under the influence of alcohol. Based on the fuel cell technology, the function is very reliable and easy to use. It operates on blowing through the wireless Alco guard if safe, it will give a ready to start signal. The accurate algorithm through the learning and reinforced decision making is making the Volvo Company thrive. Methodology The constant velocity assumption is used to model traffic vehicles, which is established upon the Kalman filter estimations of the identified vehicle. Each vehicle is given a set detection uncertainty of 2m. A quadratic curve was fit to data obtained from inaccuracies in the forward speed hypothesis aiming a range of six standard deviations, and an incremental ambiguity per sampling interval is compounded forward in time. This allows the model to account for some of the traffic vehicles' accelerations and braking (Al-Nuaimi et al., 2021). Based on the goal trajectory and three possible acceleration profiles, the automobile has comparable future predictions of its behavior. The car's prediction mistakes are lower because its intents are known ahead of time. The anticipated positioning of the car is evaluated over the anticipated location of all traffic vehicles at each time step till the automobile has completed the junction maneuver. If there is a region overlap, the activity is flagged as dangerous. Actions that have been labeled as safe are sent onto the network as allowed (Al-Nuaimi et al., 2021). The framework waits at the junction unless there are allowed activities or if the connectivity chooses to wait. Otherwise, the car accelerates to the selected speed until it reaches the desired speed. The algorithm trains in a virtual environment before testing the learnt policy on a self-driving vehicle. Data was acquired from an autonomous car in a variety of terrains, including unmarked roads and intersections, as well as diverse weather situations (Li et al., 2017). Two criteria will be used to evaluate the learnt models: the average number of instances steps a traffic vehicle brakes per trial as well as the minor axis length in between autonomous car as well as the adjacent traffic automobile per trial. Over 1000 trials, statistics will be gathered. Because both objectives can be enhanced by boosting the safety factor, the research will also run an experiment in which the proposal will enhance the safety factor of a rule-based only agent in order to ensure that the learned policy improves (Mirchevska et al.,2018). The system will make use of a probabilistic model inspector to detect safe actions and limit the RL agent's actions. The character vehicle, on the other hand, must get to the target as quickly as possible. Prospective Outcomes The system presents a novel technique that systematically estimates the maximal safe set using a fundamental technique to acquire the program's unpredictable characteristics depending on a Gaussian system design (Li et al., 2017). In comparison to current approaches, a predictive control technique derived from real model evaluation maintains safety during weaker parameters. Safety is also included in the reinforcement learning effectiveness statistic, allowing for a more seamless connection of safety and learning (Leitner et al., 2020). This system will show how the suggested technique achieves in maintaining safety where existing practices failed to avoid an unsafe condition using computations in a cart-pole system as well as an operational quad rotor implementation. References Mirchevska, B., Pek, C., Werling, M., Althoff, M., & Boedecker, J. (2018). High-level decision making for safe and reasonable autonomous lane changing using reinforcement learning. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC) (pp-). IEEE. Leitner, A., Watzenig, D., & Ibanez-Guzman, J. (2020). Validation and verification of automated systems. Springer,. Li, N., Oyler, D. W., Zhang, M., Yildiz, Y., Kolmanovsky, I., & Girard, A. R. (2017). Game theoretic modeling of driver and vehicle interactions for verification and validation of autonomous vehicle control systems. IEEE Transactions on control systems technology, 26(5),-. Schwarting, W., Alonso-Mora, J., & Rus, D. (2018). Planning and decision-making for autonomous vehicles. Annual Review of Control, Robotics, and Autonomous Systems, 1, 187-210. Al-Nuaimi, M., Wibowo, S., Qu, H., Aitken, J., & Veres, S. (2021). Hybrid Verification Technique for Decision-Making of Self-Driving Vehicles. Journal of Sensor and Actuator Networks, 10(3), 42. Iberraken, D., Adouane, L., & Denis, D. (2018). Multi-level bayesian decision-making for safe and flexible autonomous navigation in highway environment. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp-). IEEE. Selvaraj, Y., Ahrendt, W., & Fabian, M. (2019). Verification of decision making software in an autonomous vehicle: An industrial case study. In International Workshop on Formal Methods for Industrial Critical Systems (pp. 143-159). Springer, Cham.