2026· IEEE Transactions on Automation Science and Engineering· Vol 23, pp. 13220-13232· 0 citations· 34 references
Abstract
Visual localization is vital for autonomous systems but remains challenging under dynamic conditions. Transformers offer strong temporal modeling at quadratic cost, while CNNs are efficient yet limited in long-range dependencies. Existing methods also lack robustness to illumination, weather, and seasonal changes, constraining real-world applicability. To address this, this paper proposes AdapseqNet, a dual-branch architecture that integrates stabilized state-space modeling with differential temporal enhancement. First, a stabilized state-space formulation featuring Lyapunov-constrained parameterization and adaptive discretization is proposed, ensuring asymptotic stability and linear computational complexity for reliable processing of extended sequences. Second, a selective Mamba architecture is developed to combine temporal-state modeling with content-aware gating, enabling adaptive feature selection that emphasizes discriminative cues while suppressing redundancy. Third, a differential enhancement module is designed to extract motion-invariant representations through symmetric temporal differencing and LSTM-based refinement, enhancing resilience to appearance variations caused by lighting, weather, and seasonal changes. Beyond architectural design, multi-scale feature fusion and output distribution control are incorporated to optimize representation quality and ensure consistency for similarity-based retrieval. Extensive experiments on multiple benchmarks demonstrate that AdapseqNet achieves a better localization accuracy across diverse and challenging conditions. Note to Practitioners—Visual localization is crucial for autonomous robots but often fails under varying lighting, weather, or seasonal conditions. We propose a dual-path approach: one path captures long-term patterns using control-inspired stable modeling, while the other extracts motion cues that remain consistent despite appearance changes. This combination enables accurate place recognition even in extreme environments. Our system operates efficiently on standard hardware and was tested on an indoor robot, achieving centimeter-level accuracy. This approach can enhance existing navigation systems without requiring additional sensors. Future work will focus on real-time optimization for outdoor deployment.
Vehicle-to-everything (V2X) collaboration can alleviate the limited perception range and occlusion issues of single-agent autonomous driving. However, most existing cooperative studies still focus on single-frame perception, while the few works on joint cooperative perception and prediction largely rely on fixed-step, frame-aligned fusion, making them difficult to apply in realistic systems with cross-agent temporal misalignment. To address this issue, we propose CoSPACE, a novel framework that formulates asynchronous cooperative perception and prediction as continuous evolution and event-triggered correction of a shared scene state. Specifically, CoSPACE maintains an ego-centric global BEV state and propagates it to arbitrary observation timestamps using an ODE-style dynamics model, thereby preserving temporal coherence under irregular multi-agent inputs. When asynchronous observations arrive, they are treated as local evidence and assimilated by an event-triggered update module, which first selects informative regions and then performs time-aware gated correction to adaptively incorporate temporally reliable evidence. With this design, CoSPACE moves beyond the conventional frame-aligned fusion paradigm and mitigates the spatial misalignment, semantic confusion, and prediction degradation caused by temporal asynchrony. Experiments on V2XPnP-Seq under diverse delay settings show that CoSPACE consistently outperforms representative baselines and achieves strong robustness in asynchronous cooperative perception and prediction.
Hanxiao Ren, Keqiang Li, Xiang Zhao et al.· IEEE Robotics and Automation...· 0 citations
Direction-of-arrival (DoA) estimation has benefited substantially from advances in deep learning architectures. Despite the success of deep learning in static DoA estimation, most existing approaches rely on discrete snapshot processing, which complicates multi-target tracking and limits robustness under irregular sampling and signal occlusion. Conventional approaches formulate tracking as a discrete sequence regression task, which can degrade performance during target crossovers and signal occlusions. Neural ordinary differential equations have demonstrated effectiveness in modeling continuous-time dynamics and handling irregularly sampled time series; however, their application to multi-target array processing remains limited. In this paper, we propose the Factorized Physics-Informed Neural ODE (Phy-NODE), an architecture that integrates efficient static deep learning estimators with continuous-time dynamical modeling. The proposed framework factorizes the latent representation into independent state vectors, each governed by a learned differential equation. Training is performed using a tripleloss objective that combines sequence-level permutation-invariant training, motion smoothness regularization, and Bartlett beam-forming power maximization, drawing inspiration from physics-informed learning principles. The proposed method is evaluated on multi-target DoA tracking scenarios, with particular emphasis on robustness under signal occlusion conditions.
Constantinos M. Mylonakis, Nikolaos Evangelidis, Pantelis Velanas et al.· 2026 6th International Confe...· 0 citations
Trajectory prediction is essential for many robotic applications, yet most existing models rely on fixed-length observations and struggle with temporally irregular inputs. In real-world settings, prediction difficulty further increases when agents exhibit strong maneuverability, as their future motions depend on distinct short-term and long-term temporal cues. A Maneuver-aware Temporal Fusion framework is proposed to separate short-term dynamics from long-term intentions and fuse them through a motion-complexity-guided attention mechanism. The framework first extracts temporal features at different scales, and then adaptively balances them according to the maneuver patterns of each agent. To support incomplete or short observations, a self-distillation strategy is introduced to reconstruct missing motion segments, enabling consistent prediction without relying on explicit teacher-student models. Furthermore, a Mamba-Transformer hybrid backbone is employed to enhance computational efficiency and improve generalization under arbitrary observation lengths. Experiments on the ETH/UCY and SDD datasets show that MaTF consistently outperforms existing methods, particularly in scenarios with irregular or shortened observations.
Shuobo Wang, Wenyuan Qin, Yongzhao Hua et al.· IEEE Robotics and Automation...· 0 citations
Striking an optimal balance between robust tracking performance and real-time computational efficiency remains a significant challenge in the field of Unmanned Aerial Vehicle (UAV) tracking. UAV operational environments are characterized by rapid viewpoint changes, severe occlusions, and stringent onboard energy constraints, which impose rigorous demands on tracking algorithms. While recent generative trackers formulated as sequence interpretation tasks have achieved remarkable accuracy, their heavy autoregressive computations often result in high latency. Conversely, lightweight discriminative trackers frequently lack the temporal context required to handle long-term occlusions. To bridge this gap, we propose DAR-Track, a novel framework that harmonizes dynamic computation with generative modeling. First, we introduce an Adaptive Depth Gating (ADG) strategy to optimize the backbone encoder. This mechanism acts as an intelligent regulator, dynamically quantifying feature redundancy via inter-layer Euclidean distances and selectively bypassing saturated Transformer layers. Second, instead of traditional regression heads, we design a Prompt-Guided Autoregressive Decoder (PGAD). This module leverages historical coordinate sequences as spatio-temporal prompts to guide the generation of current target tokens. Furthermore, we employ a decoupled parallel attention mechanism to accelerate the decoding process. Extensive experiments on four authoritative benchmarks demonstrate that DAR-Track outperforms state-of-the-art methods, including MixFormer and SGLATrack, while maintaining superior inference speeds suitable for real-time aerial robotics.
Wenqin Dong· International Conference on...· 0 citations
Robust state estimation in GPS-denied environments remains a primary challenge for Visual-Inertial Odometry (VIO). Traditional VIO pipelines relying on hand-crafted features often experience drift or tracking failure under rapid motion and dynamic lighting. While deep learning-based local features offer improved robustness, high computational costs frequently limit their adoption in real-time, resource-constrained systems. In this work, we propose XVINS, a hybrid VIO frontend integrating XFeat—a lightweight deep feature extractor—into the optimization-based VINS-Fusion framework. The system employs a dual-strategy tracking mechanism: computationally efficient KLT optical flow is utilized for high-frequency temporal tracking, while XFeat in-ference is dynamically triggered for feature replenishment during tracking degradation. We evaluate this architecture across the 11 sequences of the EuRoC MAV dataset. The results indicate improved robustness to motion blur, reducing absolute trajectory error by up to 82.7% in highly dynamic scenarios compared to baseline methods. Furthermore, we analyze the theoretical and practical tradeoffs between deep feature quantization and classical sub-pixel precision, presenting XVINS as a viable, real-time state estimation solution for agile Micro-Aerial Vehicles (MAVs) and mobile platforms.
Thura Peou, Sarot Srang, Lychek Keo· E3S Web of Conferences· 0 citations
Continuous and reliable localization is essential for autonomous driving. Cross-view visual localization matches ground images with satellite maps, providing complementary localization cues for pipelines that depend on Global Navigation Satellite System (GNSS) signals and high-definition (HD) maps. Most existing cross-view visual localization methods process each frame independently, leaving temporal information underused and limiting accuracy under dynamic occlusion, illumination variation, and repetitive textures. This study proposes a temporal-context-enhanced framework for cross-view sequence visual localization. The proposed recurrent cross-frame module aggregates historical context from the previous state to enhance the coarse ground feature of each current frame. These enhanced features facilitate satellite candidate-region classification, while hierarchical fine-grained features enable precise local offset estimation. On the CVIS dataset, the proposed method reduces mean localization error from 3.80 m to 1.57 m and increases R@1 m from 8.14% to 40.22%. Direct transfer to KITTI-CVL achieves a mean error of 2.61 m, with target-domain fine-tuning further reducing the mean error to 2.27 m. Zero-shot field experiments on a real-world vehicle achieve a mean error of 2.84 m and R@5 m of 96.86%. These results demonstrate that temporal context enhancement significantly improves cross-view localization accuracy and supports robust deployment on public benchmarks and real-world roads.