Haowen Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Haowen Wang has a total of 28 co-authors. Here are the top co-authors by number of shared publications:
- Bocheng Zhu
School of Electronics, Peking University, Beijing 100871, China
2 shared publications - Guangchen Li
Peking University
2 shared publications - Jiangyi Qin
National Innovation Institute of Defense Technology, Academy of Military Science, Beijing, China
2 shared publications - Junren Sun
School of Artificial Intelligence, China University of Mining and Technology-Beijing, Beijing 100091, China
2 shared publications - Junyu Wei
College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China
2 shared publications - Ni Zhu
AME-GEOLOC, Gustave Eiffel University, Bouguenais, France
2 shared publications
Latest Publications
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 11 (10) , 11936-11943
AeroDiffusion: Real-Time Quadrotor Trajectory Planning in Dynamic Environments via Conditional Diffusion
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 11 (10) , 11936-11943
AeroDiffusion: Real-Time Quadrotor Trajectory Planning in Dynamic Environments via Conditional Diffusion
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (10) , 23015-23031
MC-GVIO: Motion Consistency Aided GNSS-Visual-Inertial Odometry for Land Vehicle Navigation in Urban Environments
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (10) , 23015-23031
MC-GVIO: Motion Consistency Aided GNSS-Visual-Inertial Odometry for Land Vehicle Navigation in Urban Environments
- 9 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (6) , 6647-6663
Multi-Agent Transformer Learning for Moving Target Positioning and Tracking in Complex Environments Using UAV Swarms
This paper presents a novel approach that integrates Time Difference of Arrival (TDOA)-based positioning with Multi-Agent Reinforcement Learning (MARL) to enhance the performance of Unmanned Aerial Vehicle (UAV) swarms in tracking and locating a moving ground target within complex environments. To ensure optimal target assignment, we employ the Hungarian algorithm, enabling each UAV to track a designated target and maintain a spatial formation that minimizes the geometric dilution of precision (GDOP). The coordinated behavior of the UAV swarm is then formulated as a partially observable Markov decision process (POMDP), where a Multi-Agent Transformer (MAT) model, built upon the multi-agent advantage decomposition theorem, is employed to map UAV observations to coordinated actions effectively. Finally, the UAVs adjust their trajectories based on the generated actions to improve positioning accuracy, avoid obstacles, and conserve energy. Through extensive simulations, the proposed approach demonstrates its effectiveness in continuously and accurately tracking a moving target, even in dynamic and challenging environments.
- 9 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (6) , 6647-6663
Multi-Agent Transformer Learning for Moving Target Positioning and Tracking in Complex Environments Using UAV Swarms
This paper presents a novel approach that integrates Time Difference of Arrival (TDOA)-based positioning with Multi-Agent Reinforcement Learning (MARL) to enhance the performance of Unmanned Aerial Vehicle (UAV) swarms in tracking and locating a moving ground target within complex environments. To ensure optimal target assignment, we employ the Hungarian algorithm, enabling each UAV to track a designated target and maintain a spatial formation that minimizes the geometric dilution of precision (GDOP). The coordinated behavior of the UAV swarm is then formulated as a partially observable Markov decision process (POMDP), where a Multi-Agent Transformer (MAT) model, built upon the multi-agent advantage decomposition theorem, is employed to map UAV observations to coordinated actions effectively. Finally, the UAVs adjust their trajectories based on the generated actions to improve positioning accuracy, avoid obstacles, and conserve energy. Through extensive simulations, the proposed approach demonstrates its effectiveness in continuously and accurately tracking a moving target, even in dynamic and challenging environments.
- Open Access
- 3 March 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 318
3DLaneDT: 3D lane detection via reliable depth and temporal fusion for autonomous driving
- 3 March 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 318
Open Access3DLaneDT: 3D lane detection via reliable depth and temporal fusion for autonomous driving
- 8 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 22 (15455955) , 20951-20965
SWIFT: A Distributed One-Stage Planner for Efficient Multi-Quadrotor Trajectory Optimization
This paper presents SWIFT (Swarm-Wise Inference for Fast Trajectory Planning), a distributed one-stage planner designed for efficient multi-quadrotor trajectory optimization in cluttered environments. SWIFT unifies depth-based perception, interaction-aware modeling, and trajectory prediction into a single lightweight network, enabling decentralized real-time planning without reliance on global maps. Each quadrotor processes its local depth observation and asynchronously received neighbor intentions to predict multiple candidate trajectories, leveraging a confidence-guided cross-attention mechanism to selectively model critical inter-agent interactions. The proposed framework achieves sub-millisecond planning latency while ensuring collision avoidance, dynamic feasibility, and cooperative behavior under limited communication. Furthermore, a visual-inertial odometry drift correction strategy based on inter-agent visual detections enhances positioning accuracy during real-world flights. Extensive simulations and forest experiments validate that SWIFT outperforms existing baselines in trajectory smoothness, success rate, scalability, and robustness across diverse scenarios.
- 8 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 22 (15455955) , 20951-20965
SWIFT: A Distributed One-Stage Planner for Efficient Multi-Quadrotor Trajectory Optimization
This paper presents SWIFT (Swarm-Wise Inference for Fast Trajectory Planning), a distributed one-stage planner designed for efficient multi-quadrotor trajectory optimization in cluttered environments. SWIFT unifies depth-based perception, interaction-aware modeling, and trajectory prediction into a single lightweight network, enabling decentralized real-time planning without reliance on global maps. Each quadrotor processes its local depth observation and asynchronously received neighbor intentions to predict multiple candidate trajectories, leveraging a confidence-guided cross-attention mechanism to selectively model critical inter-agent interactions. The proposed framework achieves sub-millisecond planning latency while ensuring collision avoidance, dynamic feasibility, and cooperative behavior under limited communication. Furthermore, a visual-inertial odometry drift correction strategy based on inter-agent visual detections enhances positioning accuracy during real-world flights. Extensive simulations and forest experiments validate that SWIFT outperforms existing baselines in trajectory smoothness, success rate, scalability, and robustness across diverse scenarios.
- 21 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 26 (6) , 9047-9064
Dynamic Process Noise Covariance Adjustment in GNSS/INS Integrated Navigation Using GRU-SAC for Enhanced Positioning Accuracy
The Kalman filter is widely used in GNSS/INS integrated navigation systems to fuse information, resulting in high precision and robust positioning performance. In the Kalman filter, the accuracy of the process noise covariance matrix directly affects the precision of the positioning results. We propose a Soft Actor-Critic (SAC) algorithm based on Gated Recurrent Unit neural networks (GRU-SAC) to dynamically adjust the process noise covariance matrix online using sequential observation data to improve positioning accuracy. We model the decision-making process as a Partially Observable Markov Decision Process (POMDP) and incorporate multiple information sources as the system state. The GRU network is used to extract temporal features from the navigation data and to address memory consumption issues commonly associated with POMDPs. And the SAC algorithm continuously adjusts the process noise covariance based on observations from the Kalman filter, allowing the algorithm to perform better in complex, dynamic, and changing navigation environments. Additionally, we provide detailed design and deployment strategies for both loosely-coupled and tightly-coupled systems. Extensive experiments have been conducted to validate the effectiveness of our method. The results show that our approach generalizes well across a wide range of preset process noise covariance matrices and performs excellently in suppressing error drift during GNSS outages.
- 21 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 26 (6) , 9047-9064
Dynamic Process Noise Covariance Adjustment in GNSS/INS Integrated Navigation Using GRU-SAC for Enhanced Positioning Accuracy
The Kalman filter is widely used in GNSS/INS integrated navigation systems to fuse information, resulting in high precision and robust positioning performance. In the Kalman filter, the accuracy of the process noise covariance matrix directly affects the precision of the positioning results. We propose a Soft Actor-Critic (SAC) algorithm based on Gated Recurrent Unit neural networks (GRU-SAC) to dynamically adjust the process noise covariance matrix online using sequential observation data to improve positioning accuracy. We model the decision-making process as a Partially Observable Markov Decision Process (POMDP) and incorporate multiple information sources as the system state. The GRU network is used to extract temporal features from the navigation data and to address memory consumption issues commonly associated with POMDPs. And the SAC algorithm continuously adjusts the process noise covariance based on observations from the Kalman filter, allowing the algorithm to perform better in complex, dynamic, and changing navigation environments. Additionally, we provide detailed design and deployment strategies for both loosely-coupled and tightly-coupled systems. Extensive experiments have been conducted to validate the effectiveness of our method. The results show that our approach generalizes well across a wide range of preset process noise covariance matrices and performs excellently in suppressing error drift during GNSS outages.
- 3 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 73 (11) , 16295-16311
Online Semi-Supervised Transformer for Resilient Vehicle GNSS/INS Navigation
The Inertial Navigation Systems (INS) and Global Navigation Satellite Systems(GNSS) integrated navigation system is widely employed for vehicular positioning. However, obstacles in the wireless signal path often lead to GNSS outages, diminishing the reliability of positioning. Many deep learning based methods have been proposed for their high accuracy and outstanding generalization capabilities. Nevertheless, they overlook the constraints posed by limited computational resources and memory capacity within real-time, online navigation systems. To strike a balance between positioning accuracy and model complexity, this paper proposes an Online Semi-supervised Transformer (OSS-Transformer) model for GNSS/INS integrated navigation that functions effectively during GNSS signal outages and has low computational requirements. This model comprises three critical components: (1) A Transformer-based architecture for adaptive feature extraction; (2) A self-supervised Masked Language Model (MLM) task to understand contextual relationships and feature distributions within INS data; (3) An improved memory replay technique for efficient online network training. Verified by the road test, the proposed OSS-Transformer based integrated navigation system reaches a maximum error of 6.02 m during 40 s outages and 9.81 m during 100 s outages, significantly outperforming traditional deep-learning methods. Additionally, the improved memory replay technique reduces loss by 26.23% compared to conventional methods, making it feasible on navigation devices with constrained computational resources and memory.
- 3 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 73 (11) , 16295-16311
Online Semi-Supervised Transformer for Resilient Vehicle GNSS/INS Navigation
The Inertial Navigation Systems (INS) and Global Navigation Satellite Systems(GNSS) integrated navigation system is widely employed for vehicular positioning. However, obstacles in the wireless signal path often lead to GNSS outages, diminishing the reliability of positioning. Many deep learning based methods have been proposed for their high accuracy and outstanding generalization capabilities. Nevertheless, they overlook the constraints posed by limited computational resources and memory capacity within real-time, online navigation systems. To strike a balance between positioning accuracy and model complexity, this paper proposes an Online Semi-supervised Transformer (OSS-Transformer) model for GNSS/INS integrated navigation that functions effectively during GNSS signal outages and has low computational requirements. This model comprises three critical components: (1) A Transformer-based architecture for adaptive feature extraction; (2) A self-supervised Masked Language Model (MLM) task to understand contextual relationships and feature distributions within INS data; (3) An improved memory replay technique for efficient online network training. Verified by the road test, the proposed OSS-Transformer based integrated navigation system reaches a maximum error of 6.02 m during 40 s outages and 9.81 m during 100 s outages, significantly outperforming traditional deep-learning methods. Additionally, the improved memory replay technique reduces loss by 26.23% compared to conventional methods, making it feasible on navigation devices with constrained computational resources and memory.