Deep Reinforcement Learning for Human-Like Driving Policies in Collision Avoidance Tasks of Self-Driving Cars

Abstract

The technological and scientific challenges involved in the development of autonomous vehicles (AVs) are currently of primary interest for many automobile companies and research labs. However, human-controlled vehicles are likely to remain on the roads for several decades to come and may share with AVs the traffic environments of the future. In such mixed environments, AVs should deploy human-like driving policies and negotiation skills to enable smooth traffic flow. To generate automated human-like driving policies, we introduce a model-free, deep reinforcement learning approach to imitate an experienced human driver's behavior. We study a static obstacle avoidance task on a two-lane highway road in simulation (Unity). Our control algorithm receives a stochastic feedback signal from two sources: a model-driven part, encoding simple driving rules, such as lanekeeping and speed control, and a stochastic, data-driven part, incorporating human expert knowledge from driving data. To assess the similarity between machine and human driving, we model distributions of track position and speed as Gaussian processes. We demonstrate that our approach leads to humanlike driving policies.

Multiple vehicle cooperation and collision avoidance in automated vehicles: survey and an AI-enabled conceptual framework

“…The developed agents are learned and can operate in unpredictable, broad, and stochastic contexts, as revised. The agent has been particularly trained by the effective way of a combination of Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) 170 . We proposed a multi-agent DRL based on an ideal driving strategy for avoidance or mitigating multiple collisions.…”

Implementing Autonomous Driving System in a Virtual Environment

“…Based on the action performed in the environment the agent receives rewards. If the action performed by the agent is positive, then the reward is positive, but if the action performed by the agent is negative such as hitting some other vehicle, people or driving through sidewalk the reward is negative [8][9][10][11].The learning approach has an environment, with which the agent reacts, states at which the agent might remain, actions performed by the agent which brings changes in the states of an agent. As the agent performs action, the performed action derives some rewards and based on the reward the agent changes its states in the environment [12].…”

Human-like Decision Making for Autonomous Vehicles at the Intersection Using Inverse Reinforcement Learning

“…That is, the current driverless strategy does not consider the driving behavior and decision-making intelligence, based on the long-term accumulated driving experience of humans. In the future, intelligent vehicles with autonomous driving ability will travel the roads alongside with vehicles driven by human drivers [6], in a mixed traffic scenario that will exist for a long time [1,[7][8][9]. In order to reduce the discomfort of human drivers, in the face of intelligent vehicles with automatic driving function, it is necessary for autonomous vehicles to learn the behavior and logic of human drivers.…”

“…Human driving behavior modeling mainly follows two approaches: model-based method and data-based method. Model-based methods require a priori knowledge, which can be obtained from experiments, physics, and so on, while they are usually divided into perception, decision-making and execution modules [7]. Autonomous vehicle driving basic tasks (such as steering, speed control, overtaking and obstacle avoidance) are based on model predictive control, Markov chain modeling and adaptive control.…”