In order to improve trajectory tracking accuracy, a reinforcement learning method was employed to address the trajectory tracking task in autonomous driving. There are many conveniences and advantages in theories, methods and tools utilizing reinforcement learning to solve trajectory tracking control problems. To teach an intelligent agent effective driving skills, the structured road information in the urban environment was extracted to generate an accurate reference trajectory. Then real-world scenarios were modeled to build a simulation environment for training. To effectively train the intelligent driving agent, Imitation Learning was firstly employed to teach the agent primary driving skills. Afterwards, Reinforcement Learning was adopted to optimize the agent’s driving policy. After the intelligent driving agent was well trained in the simulator, the tracking experiments were conducted in the simulator and the real-world scenarios. The proposed method was compared with base-line methods of geometric tracking, optimization-based tracking and learning-based tracking. The experimental results demonstrated that Reinforcement-Tracking can achieve accurate trajectory tracking performance and even exceed the accuracy of most baseline methods.

The advancement of communication technologies significantly enhances the capabilities for cooperative control among connected vehicles. In terms of platooning, this paper proposes a delay-dependent cloud-based nonlinear model predictive control algorithm for fuel-saving platoons which requires maintaining a desired longitudinal inter-vehicle gap. Firstly, the platoon is modeled with longitudinal dynamics, considering the constraints of physical limitations and rear-end collision avoidance based on information topology. For random communication delay compensation, a delay-dependent control algorithm is designed and combined with a cloud control strategy to achieve platoon performance and improve fuel economy. This paper conducts a comprehensive analysis of the proposed method's asymptotic stability and string stability, employing Lyapunov techniques for validation. The simulation of a 10-vehicle platoon shows high tracking accuracy and quick consensus convergence as well as high fuel efficiency under the random time delay process and networked control protocol. Finally, the real-vehicle platooning test is conducted, demonstrating that the space errors are maintained within 1 m under high-speed scenarios with realistic communication delays.

Object detection is an essential task for autonomous vehicles to ensure safety. Due to the complexity of mining environments, it is difficult to detect objects accurately and robustly. To address these issues, this paper proposes a novel 3D LiDAR-based object detection method fusing semantic and geometric features for autonomous trucks in mining environments (3DSG). A road region extraction method is presented by establishing a semantic segmentation network with a region searching strategy to eliminate off-road point clouds. To deal with the complexity of unstructured road ground point-cloud segmentation, we propose a cascaded ground detection algorithm by performing semantic segmentation filtering and rectangular grid map filtering. A clustering method is proposed fusing adaptive distance thresholds of Euclidean clusters with semantic segmentation categories to solve the problem of the over- and undersegmentation of objects caused by the sparsity of point clouds. The performance of the proposed method is examined utilizing a real mining dataset named TG-Mine-3D. Compared with state-of-the-art methods, our method achieved higher precision of 66.39%. Moreover, for the truck and pedestrian categories, the performance of our method was significantly improved by 2.66% and 5.80%, respectively. The proposed method running at 51.35 ms achieved real-time performance.

The development of vehicle-to-vehicle (V2V) improves the cooperation efficiency of the connected autonomous vehicle (CAV) platoon. However, the failure of the network communication occurs occasionally in the realistic environment, where the ideal fixed information flow topology (FIFT) cannot be adapted. To address this issue, this paper proposes a dynamic information flow topology (DIFT) utilizing a distributed model predictive control (DMPC) algorithm for CAV platoons. When the communication link is broken, the platoon control system will switch to the corresponding collaborative control mode instead of the degeneration to adaptive cruise control (ACC). First, the duty-vehicle dynamic model is constructed. In addition, the constraints with vehicle physical limitations and rear-end collision are considered. The acceleration information of the pedal actuator from the leading vehicle and the states of the predecessor including position, velocity and acceleration are transmitted to the following vehicle with a switch Indicator under DIFT. The cost function with the consideration of DIFT and fuel consumption is formulated for the optimization problem. Comparing with the FIFT, the proposed method is evaluated in the co-simulation of Matlab-TruckSim. The results demonstrate that the proposed DIFT strategy shows the satisfactory performance of the platoon under the communication issues by measuring inter-vehicle space, position and velocity tracking, and acceleration change with high tracking accuracy of position within 1.2 m and velocity within 0.04 m/s.

Accurate extrinsic calibration of parameters between LiDAR and GPS/INS is crucial for sensor fusion and simultaneous localization and mapping in autonomous driving systems. Drift in calibrated extrinsic parameters during operation can lead to decreased accuracy and compromised safety. To address this challenge, we introduce an online calibration method for LiDAR and GPS/INS parameters utilizing multi-feature adaptive optimization. Multiple features, such as ground planes, pillars and vertices, are extracted from straight and curved trajectories, respectively. The features from adjacent frames are used to construct loss functions based on point-to-plane distance, point-to-line distance, and nearest neighbor lateral distance principles. Furthermore, we perform a theoretical analysis of the influence of different features on calibration parameters. Based on the analysis, the effective data segments are selected to avoid unnecessarily computational burden of processing all data indiscriminately. Additionally, adaptive weights are applied to combine different features to achieve high-precision calibration results. The proposed method was validated with KITTI datasets and real-world open-pit coal mine datasets, and compared with state-of-the-art (SOTA) methods. Experimental results indicate that our method effectively corrects the extrinsic calibration between LiDAR and GPS/INS online, achieving an average relative rotation error (RRE) and relative translation error (RTE) improved by 6.4% and 8.5%, respectively, compared to SOTA approaches. Moreover, the optimization time of our method is less than 13 seconds, with a maximum time efficiency improvement of 25.5% compared to the SOTA methods.

Effective temporal feature fusion is essential for improving detection accuracy in image-based bird’s-eye view perception tasks. However, achieving robust temporal feature fusion faces significant challenges due to feature misalignment caused by calibration errors and localization drift, and inaccuracies in depth estimation, all of which can lead to degraded detection performance. To overcome these challenges, we propose ACF4D, a novel temporal fusion framework designed for multi-view 3D object detection. We introduce a coarse-to-fine temporal feature alignment module aimed at mitigating temporal misalignment. Additionally, to improve the accuracy of depth estimation, we propose a temporal feature consistency loss, which constrains the discrepancies between features across aligned frames. Furthermore, we construct a temporal fusion network utilizing depthwise separable convolution to achieve efficient multi-frame feature fusion while maintaining real-time performance. Experimental results on the nuScenes benchmark demonstrate that our proposed method achieves 51.5% Mean Average Precision and 60.2% nuScenes detection score. Moreover, the model operates at 20.4 FPS, demonstrating an effective balance between detection accuracy and computational efficiency.