人机混合智能增强的自动驾驶

      针对自进化学习型算法的安全性、学习速度、学习性能等问题，提出人机混合智能增强方法，利用人类演示数据、人类在线反馈，以及人类先验知识对自学习算法的预训练、奖励函数、探索质量、安全性进行增强，实现了安全、高效、拟人的自动驾驶自进化学习能力。

       团队成员：袁康、刘涛、赵治玮、唐昕月、谷宇霄、李尚文、陈龙平、唐子豪

       相关工作：

• Evolutionary Decision-Making and Planning for Autonomous Driving based on Safe and Rational Exploration and Exploitation. Engineering, (2023)

      Decision-making and motion planning are extremely important in autonomous driving to ensure safe driving in a real-world environment. This study proposes an online evolutionary decision-making and motion planning framework for autonomous driving based on a hybrid  data- and model-driven method. First, a data-driven decision-making module based on deep reinforcement learning (DRL) is developed to pursue a rational driving performance as much as possible. Then, model predictive control (MPC) is employed to execute both longitudinal and lateral motion planning tasks. Multiple constraints are defined according to the vehicle’s physical limit to meet the driving task requirements. Finally, two principles of safety and rationality for the self-evolution of autonomous driving are proposed. A motion envelope is established and embedded into a rational exploration and exploitation scheme, which filters out unreasonable experiences by masking unsafe actions so as to collect highquality training data for the DRL agent. Experiments with a high-fidelity vehicle model and MATLAB/Simulink co-simulation environment are conducted, and the results show that the proposed online-evolution framework is able to generate safer, more rational, and more efficient driving action in a real-world environment.

• Triboelectric Nanogenerator Sensors for Intelligent Steering Wheel aiming at Automated Driving. Nano Energy, (2023): 108575.

     This paper reports a novel intelligent steering wheel developed based on the concept of triboelectricity aiming at automated driving to reduce traffic accidents. A sandwich-type sensor is designed to be integrated into the steering wheel with the aim of identifying driver’s steering intention. The steering wheel of a vehicle is furnished with a triboelectric nanogenerator (TENG)-based sensor for detecting driver intention. The superiority of the TENG-based sensor is demonstrated by comparing it to other available sensors within a vehicle. By employing different machine learning techniques, we develop classification models based on driving data from multiple drivers. We show that the faster reaction time of the TENGbased sensor can aid in emergency obstacle avoidance when compared to the regular steering wheel sensor through the use of model-predictive control. The fusion of data generated by the proposed TENG-based sensor and advanced control model represents a crucial step towards the development of an intelligent steering wheel for automated systems. This will improve the human-machine interaction for vehicle control, ultimately resulting in more efficient and effective control of the vehicle.

• Evolutionary Decision-Making and Planning for Autonomous Driving: A Hybrid Augmented Intelligence Framework (Under Review)

      Recently, thanks to the introduction of human feedback, Chat Generative Pre-trained Transformer (ChatGPT) has achieved remarkable success in the language processing field. Analogically, human drivers are expected to have great potential in improving the performance of autonomous driving under real-world traffic. Therefore, this study proposes a novel framework for evolutionary decision-making and planning by developing a hybrid augmented intelligence (HAI) method to introduce human feedback into the learning process. In the framework, a decision-making scheme based on interactive reinforcement learning (Int-RL) is first developed. Specifically, a human driver evaluates the learning level of the ego vehicle in real-time and intervenes to assist the learning of the vehicle with a conditional sampling mechanism, which encourages the vehicle to pursue human preferences and punishes the bad experience of conflicts with the human. Then, the longitudinal and lateral motion planning tasks are performed utilizing model predictive control (MPC), respectively. The multiple constraints from the vehicle's physical limitation and driving task requirements are elaborated. Finally, a safety guarantee mechanism is proposed to ensure the safety of the HAI system. Specifically, a safe driving envelope is established, and a safe exploration/exploitation logic based on the trial-and-error on the desired decision is designed. Simulation with a high-fidelity vehicle model is conducted, and results show the proposed framework can realize an efficient, reliable, and safe evolution to pursue higher traffic efficiency of the ego vehicle in both multi-lane and congested ramp scenarios.

• Feedback is all you need: from ChatGPT to autonomous driving. Science China Information Sciences, 2023, 66(6): 1-3.

• Pedestrian Collision Risk Assessment Based on State Estimation and Motion Prediction. IEEE Transactions on Vehicular Technology, 2021, 71(1): 98-111

      Active pedestrian collision avoidance (APCA) systems can significantly reduce road injuries and thus have attracted considerable attention from both the automobile and the transportation industries. Numerous studies have focused on APCA; however, it remains challenging to model variable and complex scenarios in a safe, efficient and low-cost way. For this purpose, this paper proposes a novel multipedestrian collision risk assessment framework comprising a motion prediction module, a collision checking module and a collision risk assessment module. First, the motion of the ego vehicle (EV) is predicted through a constant-acceleration (V-CA) model and a constant-turn-rate and constant-velocity (V-CTRV) model in different scenarios. Next, the pedestrian motion is estimated using Kalman filter (KF) approaches based on a constant-velocity (PCV) model and a constant-acceleration (P-CA) model. Then, the potential collision area (PCA) is defined according to the predicted motions of the EV and pedestrian, and the time-to collision (TTC) is selected to conduct collision checking. Finally, the most dangerous pedestrian is identified as the target with the minimum TTC in a multipedestrian scenario, and field tests are conducted on an autonomous vehicle platform. Comparative simulations indicate that the P-CA-based KF is more accurate and robust in variable-velocity scenarios than the P-CV-based KF. Furthermore, multipedestrian scenario simulations validate the effectiveness of the proposed framework.  

• Personalized Decision-making and Control for Automated Vehicles based on Generative Adversarial Imitation Learning. (Under Review)

      Automated driving is one of the main trends in automotive field. However, existing algorithms for automated driving are mostly designed based on statistical rules without considering the onboard drivers’ individual features, which may lead to human-vehicle inconsistency during driving. In order to meet the driving preferences of any individual driver rather than classify into several driving styles (e.g. aggressive, conservative, moderate), this paper proposes a policy-learning method based on generative adversarial imitation learning. Specifically, an imitation learning framework consisting of generators and discriminators is built to train a policy network. Then, expert models are established to acquire driving data based on characteristics of individual drivers. In addition, the feasibility of the method is verified under scenarios of car following and lane changing. Finally, the generalization ability of the method is verified compared to the behavioral cloning method. The results reveal that the proposed method is capable to mimic the personalized features of human drivers.

• Mandatory Lane-Changing Decision-Making in Dense Traffic for Autonomous Vehicles based on Deep Reinforcement Learning. 2022 6th CAA International Conference on Vehicular Control and Intelligence (CVCI). 2022.

      Mandatory lane changing in complex and crowded traffic environment is a great challenge for autonomous vehicles. This paper proposes a decision-making model based on reinforcement learning, encoded prior knowledge and traffic rules for exiting tasks within a limited distance. First, a decision policy of acceleration and lane-change actions is constructed based on the reinforcement learning technique. Also, driver prior knowledge and traffic rules are encoded to construct behavior constraints, which helps the agent explore efficiently and save training time. In addition, in order to dynamically adjust the lane-change style, the lane-change urgency that varies with the remaining distance is introduced into the framework. A challenging highway exiting scenario is designed to verify the performance of the proposed method. The results indicate that the proposed model can change the driving style according to task urgency, switching from actively overtaking and changing lanes to passively waiting for a safety gap. Compared with the traditional rule-based method, it performs better in the highway exiting task under high traffic density and limited remaining distance with a greater success rate of more than 95%.

• Planning and Control of Autonomous Driving in Lane-change Maneuver based on MPC: A Framework and Design Principles (Under Review)

      To realize a safe trajectory planning and control performance based on Model Predictive Control (MPC) and investigate the effects of MPC horizons on the system performance, this paper proposes a longitudinal-lateral decoupled trajectory planning method with a safety guarantee principle and develops an MPC tracking controller using the vehicle dynamics model. Besides, the effects of the planning horizon and the control horizon on lane-changing performance are investigated. Finally, a comprehensive evaluation indicator considering safety, comfort, and efficiency is proposed to provide recommended prediction horizons for different styles of driving behavior. Several cases are studied and the results validate the effectiveness of the proposed method.

机理数据混合增强的自动驾驶

      针对当前自进化算法安全性差、学习效率低等问题，提出机理数据混合增强的自动驾驶安全自进化方法，实现自动驾驶算法性能的安全稳步提升。

       团队成员：杨硕、王俐文、王曹俊、李时珍、张渲潼、马振禹、张昊博、戎彧

       相关工作：

• An Efficient Self-evolution Method of Autonomous Driving for Any Given Algorithm. IEEE Transactions on Intelligent Transportation Systems, 2023.

      Autonomous vehicles are expected to achieve selfevolution  in the real-world environment to gradually cover more  complex and changing scenarios.  Reinforcement learning focuses  on how agents act in the environment to maximize the cumulative  reward, with a great potential to achieve self-evolution ability.  However, most of reinforcement learning algorithms suffer from a  low sample efficiency, which greatly limits their application in autonomous  driving.  This paper presents an efficient self-evolution  method for any given algorithm based on the combination of Soft Actor Critic (SAC) and Behavioral Cloning(BC).  First, the  states of the sample trajectory in the replay buffer are separated  and input into the given algorithm (algorithm with fundamental  performance) to get the output label of actions such that the SAC algorithm can be guided using BC to achieve fast iteration  in the direction of optimization with existing basic performance.  Then, the value iteration algorithm is combined to achieve the  proportion allocation of mixed gradient feedback, in order to  trade off exploitation and exploration.  In addition, the proposed  methodology is evaluated in simulation environment taking automated  speed control as an example.  Experiment results show that  compared with SAC algorithm, the proposed method can realize  more than three times of convergence efficiency improvement,  while without destroying the exploration enhancement advantage  of reinforcement learning algorithm, that is, the performance is  improved by 20% compared with the given algorithm (Intelligent Driver Model, IDM).  The proposed method can easily extended to  improve any given model no matter it is model-based or learningbased  algorithm.

• A Safe and Efficient Self-evolving Algorithm for Decision-making and Control of Autonomous Driving Systems (Under Review)

      Autonomous vehicles with a self-evolving ability are  expected to cope with unknown scenarios in the real-world  environment.  Take the advantage of trial and error mechanism,  reinforcement learning (RL) is able to self evolve by learning  the optimal policy, and it is particularly well suitable for solving  decision-making problems.  However, RL suffers from the weak  safety and low learning efficiency, especially in the continuous  action space.  Therefore, this paper proposes a hybrid augmented  approach called Mechanism-Experience-RL to address the aforementioned  problems.  In this approach, the definition of driving  tendency in analogy with human driving experience is proposed  to reduce the search space of the autonomous driving problem to  enhance learning efficiency, while the constrained optimization  problem based on a mechanistic model is designed to ensure  safety during the self-evolving process.  Experimental results  show that the proposed method is capable of generating safe  and reasonable actions in various complex scenarios, thereby  improving the performance of the autonomous driving system  while ensuring safety.  Compared to conventional RL, the safety  and efficiency of the proposed algorithm are greatly improved.  The training process is collision-free, and the training time is  equivalent to less than ten minutes in the real world.

• How to Guarantee Driving Safety for Autonomous Vehicles in Real-world Environment: a Perspective  on Self-evolution Mechanism (Under Review)

      A succession of accidents shows that production  vehicles with autonomous driving systems do not work safely in  real-world environment especially when face unseen scenarios.  Therefore, how to ensure autonomous systems drive more safely  becomes a challenge.  Thanks to the self-learning ability of  human-being, human drivers can gradually learn how to drive  from the driving test with typical and finite scenarios to the  real world with infinite ones.  Analogically, it is believed that the  accidents can be largely reduced once the designed autonomous  vehicles are endowed with a self-learning ability to adapt to the  unseen then to infinite scenarios in the real world.  Accordingly,  this work proposes a principle to design autonomous systems  with self-evolution feature not just for a single vehicle but for  a group.  In addition, it describes our development of a selfevolution  autonomous system as an illustrative case study of  implementing such principles in practice.  The ultimate aim is  to propose a feasible solution to speed up design process of a  fully safe autonomous system.

• 基于时序差分学习模型预测控制的 一体化自动驾驶换道策略 (Under Review)

      具有自进化能力的自动驾驶换道策略有望在复杂开放的交通环境中提升性能，以应对更多的未知场景。时序差分学习 模型预测控制(Temporal Difference Learning for Model Predictive Control, TD-MPC) 结合了有模型和无模型强化学习方法的优 势，具有学习效率高，性能优异的特点。基于此，为了提高自动驾驶换道策略的整体性能，提出基于TD-MPC 的自动驾驶 一体化换道策略。具体来说，针对自动换道问题，提出基于驾驶倾向网络的一体化自动驾驶换道策略架构，构建强化学习问 题并设计完备的奖励函数，对决策规划优化问题进行统一求解。应用TD-MPC 算法设计内部模型来预测未来状态和奖励， 实现短时域内的局部轨迹优化，同时使用时序差分学习实现对长期汇报的估计，以得到驾驶倾向网络参数。所提出方法在高 保真仿真环境中被验证，结果表明，所提出方法相比规则方案保证了行驶效率，并且提高安全性和舒适性。同时与软演员- 评论家算法(Soft Actor Critic, SAC)相比，实现了7~9 倍的学习效率提升。

可解释性人工智能与自动驾驶

      针对人工智能（AI）模型的黑箱问题，提出基于Shapley值的模型可解释方法来洞察AI模型在设计、测试和验证过程中的预测，以发现模型风险，并约束模型行为。提高了AI模型的透明度、安全性和可信度。

       团队成员：李蒙、崔志浩

       相关工作：

• Explaining a Machine-Learning Lane Change Model With Maximum Entropy Shapley Values. IEEE Transactions on Intelligent Vehicles, 2023.

      Artificial intelligence (AI) techniques have been widely implemented in the domain of autonomous vehicles (AVs). However, existing AI techniques, such as deep learning and ensemble learning, have been criticized for their black-box nature. Explainable AI is an effective methodology to understand the black box and build public trust in AVs. In this article, a maximum entropy-based Shapley Additive exPlanation (SHAP) is proposed for explaining lane change (LC) decision. Specifically, we first build an LC decision model with high accuracy using eXtreme Gradient Boosting. Then, to explain the model, a modified SHAP method is proposed by introducing a maximum entropy base value. The core of this method is to determine the base value of the LC decision model using the maximum entropy principle, which provides an explanation more consistent with the human intuition. This is because it brings two properties: 1) maximum entropy has a clear physical meaning that quantifies a decision from chaos to certainty, and 2) the sum of the explanations is always isotropic and positive. Furthermore, we develop exhaustive statistical analysis and visualization to present intuitive explanations of the LC decision model. Based on the explanation results, we attribute the causes of predictions with wrong results to model defects or sample sparsity, which provides guidance to users for model optimization.

• Shapley Value: From Cooperative Game to Explainable Artificial Intelligence.(Under Review)

      With the tremendous success of machine learning (ML), concerns about their black-box nature have grown. The issue of interpretability affects trust in ML systems and raises ethical concerns such as algorithmic bias. In recent years, the feature attribution explanation method based on Shapley value has become the mainstream explainable artificial intelligence approach for explaining ML models. This paper provides a comprehensive overview of Shapley value-based attribution methods. We begin by outlining the foundational theory of Shapley value rooted in cooperative game theory and discussing its desirable properties. To enhance comprehension and aid in identifying relevant algorithms, we propose a comprehensive classification framework for existing Shapley value-based feature attribution methods from three dimensions:  Shapley value type, feature replacement method, and approximation method. Furthermore, we emphasize the practical application of the Shapley value at different stages of ML model development, encompassing pre-modeling, modeling, and post-modeling phases. Finally, this work summarizes the limitations associated with the Shapley value and discusses potential directions for future research

自动驾驶闭环自进化技术

      针对自动驾驶安全关键场景难以遍历和穷尽，以及自动驾驶算法难以充分利用场景数据、兼容不同场景策略的问题，提出结合强化学习对抗场景生成、基于生成模型的场景特征提取和算法性能评价与安全边界分析、场景离线数据的持续学习等技术的闭环自进化方法，实现安全关键场景的高效生成，以及场景驱动的算法自进化。

       团队成员：李欣城、王兆一、陈诗阳、孙恒阳

       相关工作：

• A Survey on Self-evolving Autonomous Driving: a Perspective on Data Closed-Loop Technology. IEEE Transactions on Intelligent Vehicles, 2023.

      Self evolution refers to the ability of a system to evolve autonomously towards a better performance, which is a potential trend for autonomous driving systems based on self-learning approaches. However, current algorithms for autonomous driving still lack of self-evolving mechanisms and the capability of maintaining continuously performance-enhancing. Some recent studies turn to the data closed-loop (DCL) architecture to realize self evolution. Therefore, this study analyzes some relevant technologies and then proposes a novel design mechanism to guarantee the self-evolving performance for autonomous driving systems. Although existing data closed-loop platforms are not yet mature enough to fully achieve this purpose, it has the potential to incorporate cutting-edge technologies that will enhance their functionality. Moreover, we give some suggestions for its future directions for self-evolving autonomous driving, including some more cutting-edge technologies that can be incorporated into the DCL architecture. 

• Exploring Safety Boundaries of Autonomous Driving Systems: A Safety-Critical Scenario Generation and Analysis Framework.  (Under Review)

      It is important for an autonomous driving system to know its safety capacity no matter in design or operation phase. To achieve this, safety-critical scenarios, essential for safety evaluation, must be studied. Therefore, this paper proposes a framework to generate such scenarios and analyze the safety boundaries based on the given algorithms. Firstly, a generation method for fine-grained adversarial scenario based on reinforcement learning is proposed to achieve efficient exploration. Then, scenario features are non-linearly extracted and clustered for safety analysis. In addition, based on the distribution map of scenario features, the safety boundary of the autonomous driving algorithm is defined and analyzed. Finally, two given algorithms are evaluated with the proposed framework. The main accident categories of given algorithms are summarized B10from the perspective of vehicle interaction relationships, and their safety boundaries are quantitatively described and analyzed in conjunction with the distribution map of scenario features. The case study demonstrates the effectiveness of our proposed framework for safety-critical scenario exploration and safety analysis. The proposed framework allows for a comprehensive description of safety boundaries for any given algorithm, which is of great significance for the evaluation of autonomous vehicles.

小样本下持续学习型自动驾驶

      针对自动驾驶边缘场景的复杂性、未知性和多变性，以及人工神经网络在灾难性遗忘方面的挑战，提出任务边界自适应识别方法、非平衡持续学习能力增强的持续强化学习算法，以及跨场景的小样本元学习方法，从而构建自进化型自动驾驶系统，通过有限样本实现对新场景的持续学习。

       团队成员：邢家铭、韦登伟、崔艺馨、杜昊洋

       相关工作：

• Continual Reinforcement Learning for Autonomous Driving with Application on Velocity Control under Various Environment.  (Under Review)

      Reinforcement learning based methods are extensively studied in autonomous driving. However, most existing methods, suffering from catastrophic forgetting, only work properly under certain scenarios. Therefore, this paper proposes a continual reinforcement learning approach to improve the control adaptability of autonomous driving systems. Specifically, the framework is based on a Soft Actor-Critic algorithm, which includes a shared feature extractor with regularization loss to learn task-specific output mapping. A linear multi-policy heads structure is designed to continuously learn different tasks without interference. Velocity control under various environment is taken as a case study, and three velocity control tasks with a high degree of overlap in the observation spaces are designed. In addition, multi-modal optimization objectives considering safety, efficiency and comfort for the three tasks are designed to evaluate the effectiveness of the proposed approach. The simulation studies are conducted in CARLA, which demonstrates the efficiency of the proposed method in terms of improving the continual learning capability of the model. Overall, the proposed continual reinforcement learning framework contributes to the development of adaptive autonomous driving systems to a large extent.

• A Comprehensive Study on Self-Learning Methods and Implications to Autonomous Driving.  (Under Review)

      As artificial intelligence has already seen numerous successful applications, the upcoming challenge lies in how to realize artificial general intelligence. Self-learning algorithms can autonomously acquire knowledge and adapt to new, demanding applications, recognized as one of the most effective techniques to overcome this challenge. Although many related studies have been conducting, there is still no comprehensive and systematic review available, nor well-founded recommendations for the application to autonomous intelligent systems, especially autonomous driving. As a result, this paper comprehensively analyzes and classifies self-learning algorithms into three categories: broad self-learning, narrow self-learning, and limited self-learning. These categories are used to describe the popular usage, the most promising techniques, and the current status of hybridization with self-supervised learning, respectively. Then, the narrow self-learning is divided into three parts based on the self-learning realization path: sample self-learning, model self-learning and self-learning architecture. For each method, this paper discusses in detail its self-learning capacity, challenges, and applications to autonomous driving. Finally, the future research directions of self-learning algorithms are pointed out. It is expected that this study has the potential to eventually contribute to revolutionizing autonomous driving technology.

车路协同下的多车规控与协同进化

      针对不同场景多车协同规划控制问题，提出了利用车路协同技术的多车调度规划方法与基于个性化联邦学习的模型融合方法，实现多车安全高效通行与协同进化。

       团队成员：戢杨杰、王祥、周泽玮、周上航、杨紫茹、张馨雨

       相关工作：

• A comprehensive study of speed prediction in transportation system: From vehicle to traffic. Iscience, 2022.

      In the intelligent transportation system (ITS), speed prediction plays a significant role in supporting vehicle routing and traffic guidance. Recently, a considerable amount of research has been devoted to a single-level (e.g., traffic or vehicle) prediction. However, a systematic review of speed prediction in and between different levels is still missing. In this article, existing research is comprehensively analyzed and divided into three levels, i.e. macro traffic, micro vehicles, and meso lane. In addition, this article summarizes the influencing factors and reviews the prediction methods based on how those methods utilize the available information to meet the challenges of the prediction at different levels. This is followed by a summary of evaluation metrics, public datasets, and open-source codes. Finally, future directions in this field are discussed to inspire and guide readers. This article aims to draw a complete picture of speed prediction and promote the development of ITS.

• A Reliable Path Planning Method for Lane Change Based on Hybrid PSO-IACO Algorithm.  2021 6th International Conference on Transportation Information and Safety (ICTIS). IEEE, 2021: 1253-1258.

      The real-time performance of path planning algorithms and path continuity are crucial to motion planning. Thus, B-spline-based path planners have attracted extensive interest because of control flexibility and continuous curvature. However, the B-spline-based planning requires lots of resources to solve due to the multiple nonlinear constraints. Therefore, a new hybrid algorithm is proposed, which utilizes the comple- mentary advantages of particle swarm optimization (PSO) and improved ant colony optimization (IACO), called PSO-IACO. The proposed algorithm comprises two phases. First, the PSO ensures a fast convergence to a series of feasible rough paths, which are used to initialize the pheromone allocation and the position of IACO. Then, the IACO with the advantage of positive feedback help improves the quality of the path. Moreover, the main improvement of IACO from ACO is the pheromone updatestrategy considering the local and global search experience, which is inspired by the idea of PSO and Max-Min ant system. Simulation demonstrates that the path quality of PSO-IACO outperforms that of PSO, IACO, Midaco, and genetic algorithm(GA). It also outperforms that of Enumeration in most scenarios.The success solution rate is improved two times as compared to Midaco for some scenarios. And the execution time is reducedto 74% in comparison with Enumeration for the large-scalescenario.

• A method of Speed Prediction Based on Markov Chain Theory Using Actual Driving Cycle.  SAE Technical Paper, 2022

      As a prerequisite for energy management of hybrid vehicles, the results of speed prediction can optimize the performance of vehicles and improve fuel efficiency. Energy management strategies are usually developed based on standard driving cycles, which are too generalized to show the variability of driving conditions in different time and locations. Therefore, this paper constructs a representative driving cycle based on driving data of the corresponding time and location, used as historical information for prediction. We propose a method to construct the driving cycle based on Markov chain theory before constructing the prediction model. In this paper, multiple prediction methods are compared with traditional parametric methods. The difference in prediction accuracy between multiple prediction methods under the single time scale and multiple time scale were compared, which further verified the advantages of the speed prediction method based on Markov chain theory.

• Optimization of Roadside Sensors Placement for Cooperative Vehicle-Infrastructure System.  (Under Review)

      Cooperative vehicle-infrastructure system (CVIS), one of the key development directions of intelligent transportation, can provide autonomous vehicles with perception information beyond the visual range through roadside sensors. Optimizing sensor placement is essential to use as few sensors as possible while meeting coverage requirements. This paper establishes the coverage model of roadside sensors (including cameras, millimetre wave radars and lidars), proposes an optimization method for roadside sensor placement, and applies it to real road scenes. Firstly, the relationship between roadside sensors’ placement and coverage area is modeled according to the perceived characteristics and mounting parameters of different sensors. Then, this paper develops an optimization model for sensors placement based on the location relationship between different sensors, and employs a multi-objective Grey Wolf algorithm to solve the model. Finally, a method for optimal placement of roadside sensors using this optimization model was applied to a road section in the urban. The results show that the proposed method can use fewer sensors under the premise of satisfying the use effect, thereby saving costs and reducing energy consumption. This research contributes to realizing roadside sensors placement with high perceptual quality and developing autonomous driving.

• 多智能网联汽车轨迹规划:现状与展望.  (Under Review)

      轨迹规划是自动驾驶汽车的基本功能。随着V2X技术的发展，自动驾驶车辆具备了智能网联功能，这些汽车被称为智能网联汽车。智能网联技术可以为自动驾驶汽车带来大量信息，增强不同自动驾驶汽车之间的合作，并为轨迹规划提供额外的优化空间，以减少驾驶时间，提高驾驶舒适性和安全性。与传统的单车轨迹规划相比，多车轨迹规划可以充分利用智能网联汽车的技术优势，为多个自动驾驶汽车规划合适的轨迹。因此，本文概述了多车轨迹规划研究的现状。总结了典型的多车轨迹规划应用场景，包括结构化场景和非结构化场景。还总结了不同的多车轨迹规划合作规划策略，并分析了其特点。本文总结了用于多车轨迹规划的各种方法，包括传统的流水线规划方法和端到端方法。此外，本文对多车轨迹规划的实验也进行了归纳。最后，提出了当前多车轨迹规划面临的挑战和未来的研究方向，为智能交通系统领域的研究人员提供启发和参考。

• Towards autonomous vehicles: a survey on cooperative vehicle-infrastructure system.  (Under Review)

      Cooperative Vehicle-Infrastructure System (CVIS) is an important part of the intelligent transport system (ITS). Autonomous vehicles have the potential to improve safety, efficiency and energy saving through CVIS. Although a few CVIS studies have been conducted in the transportation field recently, a comprehensive analysis of CVIS is necessary, especially about how CVIS is applied in Autonomous vehicles. In this paper, we overview the relevant architectures and components of CVIS. After that, the state-of-the-art research and applications of CVIS in autonomous vehicles are reviewed from the perspective of improving vehicle safety, efficiency and energy saving, including scenarios such as straight road segments, intersections, ramps, etc. In addition, the datasets and simulators used in CVIS- related studies are summarized. Finally, challenges and future directions are discussed to promote the development of CVIS and provide inspiration and reference for researchers in the field of ITS.

驾驶世界模型驱动的自动驾驶

      针对经典模块化自动驾驶方法存在的误差累积、计算累积、迁移性差的问题，提出使用生成式端到端的方案解决自动驾驶问题，提出驾驶世界模型驱动的端到端方法，提取丰富的环境动力学信息，实现在城市和高速工况下可迁移且鲁棒的端到端自动驾驶。

       团队成员：杜嘉彤、白玉龙、梁越、栗烨、耿家恒、彭怀龙

       相关工作：

• A Survey on Trajectory-Prediction Methods for Autonomous Driving.  IEEE Transactions on Intelligent Vehicles, 2022, 7(3): 652-674.

      In order to drive safely in a dynamic environment, autonomous vehicles should be able to predict the future states of traffic participants nearby, especially surrounding vehicles, similar to the capability of predictive driving of human drivers. That is why researchers are devoted to the field of trajectory prediction and propose different methods. This paper is to provide a comprehensive and comparative review of trajectory-prediction methods proposed over the last two decades for autonomous driving. It starts with the problem formulation and algorithm classification. Then, the popular methods based on physics, classic machine learning, deep learning, and reinforcement learning are elaborately introduced and analyzed. Finally, this paper evaluates the performance of each kind of method and outlines potential research directions to guide readers.

• Enhanced Social Trajectory Prediction Transformer for L3 Highway Pilots.  (Under Review)

      With the rise of more and more L3 Highway Pilots, it is necessary for autonomous vehicles to make trajectory predictions of surrounding vehicles for improved driving comfort and safety. However, it remains challenging since fullcoverage HD maps in highway scenarios is very expensive, so trajectory prediction has to handle coordinate-based inputs, without HD maps. Inspired by human reasoning, we tackle these problems by using the attention mechanism and the Transformer architecture. In this article, we develop an Enhanced Social Trajectory Prediction Transformer (ESTPT) framework for L3 Highway Pilots, which adds the Social Encoder module based on the Transformer network to effectively predict the agent vehicle’s trajectories considering the interaction features without using RNN. Specifically, the three main parts of ESTPT are based on the multi-head attention mechanism, which can make accurate predictions based on coordinate-based trajectories and can deal with missing observations, to improve driving safety, comfort, and robustness. We demonstrate the performance on the NGSIM dataset and results show great improvement in terms of the positional error. Some valuable ablation analyses are conducted and the result visualization is given.

• Driving World Model: A Generative End-to-End Method for Autonomous Driving.  (Under Review)

      tEnd-to-end paradigms, which optimize the entire process in a model, have emerged as a popular trend for addressing autonomous driving challenges. However, their performance is often constrained in complex and dynamic urban environments due to limited generalization capabilities. To address this issue, this study presents the terminology of the Driving World Model (DWM), designed to simultaneously satisfies the generalizability of world representations and the robustness of control actions. Specifically, the DWM leverages a world model to extract and retain environmental dynamics information, and a brain-inspired neural controller for generating control actions, accomplishing the end-to-end autonomous driving task. The world model extracts 2D and 3D features from input images to fits the distribution of perceptual features and forecasts future feature distributions by recalling historical features. Utilizing the obtained environmental dynamics information, the braininspired neural controller is built by simulating the nematode nervous system’s operation process of perception, planning, and control, resulting in autonomous driving control outputs. Moreover, the Driving World Model can simultaneously generate BEV images of the environment to improve the interpretability of the model. We employ monocular camera data and driving data collected from the CARLA Simulator to train the proposed model, and subsequently on-policy validate its generalizability in new urban scenarios and new weather conditions. Results affirm that the model achieves state-of-the-art urban autonomous driving performance and is robust against input noise.

• FEN-DQN An End-to-End Autonomous Driving Framework Based on Reinforcement Learning with Explicit Affordance.  (Under Review)

      The slow convergence rate is a thorny problem of current end-to-end autonomous driving paradigm with various traffic elements and tasks. In this paper, we propose an end-toend autonomous driving framework FEN-DQN to simplify problems involving feature extraction network (FEN) with explicit affordance, along with some associated driving measurements such as vehicle speed and position. The FEN-DQN can be divided into two part. First, FEN is applied to map the forwardlooking camera images into explicit affordances, which represent traffic information in low-dimension. Secondly, a deep Q-network (DQN) is used to map explicit affordances to vehicle actions. Based on the CARLA simulator, we use the OpenAI-Gym to construct a simulation scenario at traffic intersections to evaluate our proposed framework. In addition, we also conduct comparative experiments on different inputs to show the excellent effect of our framework. The results showcase that FEN-DQN could converge faster and perform better compared with the other inputs at traffic intersections in the simulation scenario with the assistance of FEN.

• Quality Detection Model for Automotive Dashboard Based on an Enhanced Visual Model.  SAE Technical Paper, 2022.

      For an enterprise, product quality is the foundation of its further development. Therefore, how to detect the quality of the products produced by the assembly line and accurately identify the problematic parts has become an increasingly concerned issue for enterprises. In this paper, we propose a novel quality detection model combining the latest YOLOv5 model and convolutional neural network, which can further improve the recognition precision and accuracy of YOLOv5 on the basis of its lightweight and high recognition efficiency. The proposed model can meet the needs of complex quality problems that are difficult to detect directly in assembly-line products. In the experiment, our model can detect the automotive dashboard and judge whether the cable buckle is connected in place. The accuracy of each buckle in the picture being correctly detected is more than 98%, the classification accuracy is also expected to reach 98%.

车辆底盘动力学统一控制技术

      针对当前主动安全控制方法在传感执行异构的车辆上不具备可移植性的问题，提出了建立可重构的动力学模型以及可重构的参数辨识方法，并构建规划、运动控制和动力学控制的多目标集中式集成架构，实现智能汽车主动安全系统在传感执行异构的车辆上具备快速移植能力。

       团队成员：李云鹏、阮鑫耀、李朋、汪琦涵、闵静