Machine learning method to ensure robust decision-making of AVs

Abstract

Replacing the human driver to perform the Dynamic Driving Task (DDT)[1] will require perception, complex analysis and assessment of traffic situation. The path leading to success the deployment of fully Autonomous Vehicle (AV) depends on the resolution of a lot of challenges. Both the safety and the security aspects of AV constitute the core of regulatory compliance and technical research. The Autonomous Driving System (ADS) should be designed to ensure a safe manoeuvre and a stable behaviour despite the technological limitations, the uncertainties and hazards which characterize the real traffic conditions. In fully Autonomous Driving situation, detecting all relevant objects and agents should be sufficient to generate a warning, however the ADS requires further complex data analysis steps to quantify and improve the safety of decision making. This paper aims to improve the robustness of decision-making in order to mimic human-like decision ability. The approach is based on machine learning to identify the criticality of the dynamic situation and enabling ADS to make appropriate decision and fulfil safe manoeuvre.

A Lightweight Operation Mode Decision Method for Cleaning Robots Driven by Garbage Attributes Perception

“…Until significant advances in deep learning technology, traditional methods dominated research in the field of robotic decision control. These methods, including multilayer perceptrons, support vector machines, Bayesian networks, and AdaBoost, are widely used to solve relatively small-scale decision-control problems such as simple autonomous navigation [35], vehicle lane change decisions [36], and collision risk assessment [37]. In recent years, with significant advances in big data, deep learning models, and computational power, learning-based methods have attracted widespread research interest in the domain of robot decision control.…”

Performance Guarantees on Machine-Learning-based Overtaking Strategies for Autonomous Vehicles

“…where ∆ l,1,min , ∆ l,1,max , ∆ l,2,min , ∆ l,2,max are the boundaries of Λ l,1 , Λ l,2 . The formulations (6)- (7) show that y ref , v ref must be inside of a limited neighbourhood of the safe trajectory y i+1,s , v x,i+1,s . In practice, it is suggested to select |∆ l,i,max | = |∆ l,i,min | = ∆ l,i,m , i = {1; 2}, which leads to symmetric domains.…”

“…A reinforcement-learning-based overtaking control strategy is proposed in [5], [6]. In [7] a Q-learning strategy is used in the design of driving algorithms for multi-lane environments. An analysis method of the robust properties of the machine-learning-based overtaking decision strategies is found in [7].…”

“…In [7] a Q-learning strategy is used in the design of driving algorithms for multi-lane environments. An analysis method of the robust properties of the machine-learning-based overtaking decision strategies is found in [7]. A deep-reinforcement-learning method is used in [8], which is applied to decision making in different manoeuvres.…”

Cooperative- and Eco-Driving: Impact on Fuel Consumption for Heavy Trucks on Hills

“…In addition, machine learning algorithms are used to realize autonomous driving. Tami et al [32], for example, uses this to make decisions for lane changes. However, the presented methods mostly refer to short-term planning horizons, which are not sufficient for energy-efficient driving and, therefore, do not include eco-driving.…”