26 March 2022 in Applied Surface Science
5 May 2025 in IEEE Internet of Things Journal
Accurate prediction of vehicle motions is imperative for enabling cooperative perception and planning of autonomous vehicles, however effective modeling of complex spatio-temporal interactions and long-term dependencies between vehicles remains a formidable challenge. To tackle these issues, we propose a novel motion query-based social transformer network (MOST) for vehicle trajectory and intention prediction through a multitask approach, which is composed of temporal transformer encoder module, social interaction module and motion queries-based multitask feature decoder in a hierarchical manner. The temporal transformer is responsible for capturing long-range temporal correlations of individual motion states through a self-attention mechanism with residual connection, while the spatial interaction dependencies between vehicles are acquired through the social interaction module by constructing social tensors. Furthermore, Considering the uncertainty and diversity of future vehicle behaviors, a motion query-based feature decoder is proposed, which is equipped with learnable parameters to assimilate prior knowledge and generate multiple possible future trajectories and intentions. To assess our model’s effectiveness, we carried out comprehensive experiments on the open-source NGSIM and HighD dataset. The results demonstrated that our approach reaches unparalleled performance, with an average prediction accuracy improvement of about 50% on the NGSIM dataset and 20% on the HighD dataset compared to the state-of-the-art method.
12 January 2026 in IEEE Robotics and Automation Letters
Autonomous driving requires accurate scene understanding, including road geometry, traffic agents, and their semantic relationships. In online HD map generation scenarios, raster-based representations are well-suited to vision models but lack geometric precision, while graph-based representations retain structural detail but become unstable without precise maps. To harness the complementary strengths of both, we propose DiffSemanticFusion—a fusion framework for multimodal trajectory prediction and planning. Our approach reasons over a semantic raster–fused BEV space, enhanced by a map diffusion module that improves both the stability and expressiveness of online HD map representations. We validate our framework on two downstream tasks: trajectory prediction and planning-oriented end-to-end autonomous driving. Experiments on real-world autonomous driving benchmarks, nuScenes and NAVSIM, demonstrate improved performance over several state-of-the-art (SOTA) methods. For the prediction task on nuScenes, we integrate DiffSemanticFusion with the online HD map informed QCNet, achieving a 5.1% performance improvement. For end-to-end autonomous driving in NAVSIM, DiffSemanticFusion achieves SOTA results, with a 15% performance gain in NavHard scenarios. In addition, extensive ablation and sensitivity studies show that our map diffusion module can be seamlessly integrated into other vector-based approaches to enhance performance.
23 July 2025 in IEEE Robotics and Automation Letters
The realization of autonomous driving(AV) is closely linked to the development of intelligent decision-making modules that can operate safely in dynamic, uncertain environments. To address issues such as delayed response and poor coupling in highway scenarios, this letter proposes a hierarchical Coupled Decision-Making (CDM) framework. The CDM framework separates high-level intention planning from low-level behavior control. The upper layer uses DE-D3QN, enhanced with episodic memory buffer (EMB) and experience replay buffer (ERB), to improve learning efficiency in large-scale state spaces. The lower layer employs BF-TD3 with a barrier function to generate continuous, risk-aware control actions. Additionally, a probabilistic risk model combining lane speed gain, risk indicators, and Bayesian estimation enables adaptive evaluation of surrounding vehicle impact. To verify the contribution of each component, ablation studies are conducted on the virtual distance model and barrier function. Results show that CDM achieves better training stability, decision robustness, and a more favorable balance between risk and efficiency compared to baselines, while aligning decisions with driving expectations.
12 March 2026 in IEEE Transactions on Artificial Intelligence
The motion prediction of surrounding vehicles is crucial for autonomous driving. As it can effectively guide the planning and decision-making of the autonomous vehicle. Although many methods have been proposed to predict the future motion of vehicles, but they failed capture the essense of multitask learning for vehicle motion prediction and lacked effective semantic and interaction information fusion. In this paper, we propose a novel model, Graph Attention Transformer with Multi- Task Learning (GATran), to predict the future motion of the target vehicle by modeling motion prediction task as a multi-task learning problem of trajectory prediction and driving intention estimation. The model consists of three modules: Feature Fusion, Spatio-Temporal Interaction, and Motion Decoder. The Feature Fusion module is used to effectively fuse the semantic and dynamic information of the target vehicle. The Spatial Interaction module is used to model the interaction between vehicles. The Motion Decoder module decouples the motion prediction into multi-task learning for trajectory prediction and driving intention estimation in a parallel manner. Extensive experiments on real-world datasets, with average improvement of 50% in NGSIM and 40% in HighD, demonstrate that the proposed model achieves superior performance.
1 July 2026 in IEEE Robotics and Automation Letters