In the context of the Industrial Internet of Things (IIoT), whole-body mobile manipulation plays a critical role in smart manufacturing and flexible automation, which requires high-quality teleoperation data and reliable long-horizon, multimodal action prediction. We present an exoskeleton-VR teleoperation system with simple force feedback for single-operator whole-body mobile control, enabling consistent multimodal demonstrations. We further introduce MiMoCo, an encoder–decoder imitation learning framework with Efficient Context Modeling Network (ECM-Net) for linear-complexity phase-level temporal modeling and Multi-Receptive Field Fusion Network (MRF-Net) for dual-path multi-receptive-field visual-motor fusion. Real-world experiments on a mobile robot demonstrate that MiMoCo achieves competitive performance compared with state-of-the-art baselines across multiple whole-body mobile manipulation tasks, and excels in long-horizon, fine-grained subtasks. Source code is publicly available at https://github.com/meijie-jesse/MiMoCo.

Imitation learning has emerged as a promising paradigm for automating robotic fine manipulation, yet maintaining temporal coherence and execution precision in long-horizon tasks remains a formidable challenge. This requires addressing exponential error accumulation in long-horizon tasks, as well as the inherent trade-off between millimeter-level positioning precision and motion smoothness, which is critical for the stability and accuracy of robotic automation systems. To this end, we propose FreMotion, a task-driven asymmetric encoder-decoder framework designed to enhance the robustness and efficiency of robotic manipulation. Our approach introduces the Temporal Dynamic Encoder (TDE-Net), which employs linear attention with dynamic gating to capture long-range dependencies and mitigate error accumulation in continuous control. Simultaneously, the Frequency-Aware Decoder (FAD-Net) leverages a dual-path attention mechanism to explicitly disentangle and fuse high- and low-frequency information for precise visuo-motor coordination. Extensive experiments on bimanual robotic tasks demonstrate that FreMotion, with orders of magnitude fewer parameters (approx. 100M vs. Billions), not only outperforms state-of-the-art Vision-Language-Action (VLA) methods in success rate but also achieves faster inference time, thus ensuring high-frequency continuous control required for industrial automation.

Predicting potential collision events is beneficial to ensure the driving safety of autonomous vehicles. Existing graph-based collision prediction methods rely heavily on domain knowledge and predefined semantic relations, limiting their flexibility and adaptability in complex driving scenarios. To overcome these challenges, this paper introduces a novel collision prediction framework named HGAT-CP, which integrates a Heterogeneous Graph Attention Network (HGAT) with a Long Short-Term Memory network (LSTM) to model the spatial-temporal interactions in scenes. First, the proposed method employs a data-driven scene graph embedding module to autonomously learn relationships between vehicles and lanes and construct flexible scene graphs. Then, the HGAT module utilizes a dual-level attention mechanism, operating at both the node level and type level, to capture spatial interactions without relying on predefined semantic rules. The LSTM module models temporal dependencies of the scene graph embeddings to improve the prediction of collision events over time. Experimental evaluations on public datasets demonstrate that our proposed method achieves state-of-the-art performance, outperforming existing methods across all metrics.

Multi-view image-based BEV (Bird's Eye View) 3D perception is gaining attention as an alternative to highcost LiDAR systems and has achieved notable success. However, there is a significant safety concern for future image-based BEV autonomous driving in low-light conditions (such as nighttime) while the limited research on BEV detectors for these scenes. In this paper, we attempt to enhance low-light BEV perception with illumination-guided feature fusion. We propose Retinex-BEVFormer, which uses illumination information generated by the Retinex theory to enhance the model's robustness to varying lighting conditions and improve detection performance in low-light scenes. Additionally, to address the illumination estimation discontinuity from multi-view images that can adversely affect detection, we propose the MVB-Retinex module, which balances illumination estimation by leveraging overlapping regions between adjacent images. Notably, our proposed method is a plug-and-play module that can be applied to any image-based BEV detector method and does not require any additional ground truth supervision. We conduct extensive experiments on the nuScenes dataset, validating our algorithm in nighttime and daytime scenes. Compared to the baseline, our algorithm achieves a 2.9% increase in mAP on the validation set with minimal computational cost, especially showing a 3.6% improvement in the nighttime scene. The experiments demonstrate that our Retinex-BEVFormer effectively improves detection performance under low light conditions and enhances performance under normal illumination, indicating increased robustness of the BEV detector.

In this paper, a vehicle Re-ID framework which integrates image processing and traffic flow model is developed. First, the CNN network is applied for vehicle detection and tracking and extracting attribute recognition. Particularly, attributes including vehicle color, type, make, and Re-ID feature are extracted to derive a similarity matrix between upstream and downstream vehicles. However, solely using these features could not achieve satisfaction matching accuracy. Our testing only shows a moderate accurate of around 72.3%. To further improve the Re-ID rate, this paper integrates visual information with the well-known IDM car-following mode. In our framework, IDM is first used to estimate the arrival time window for each upstream vehicle; and then with this time window derive a filter matrix which set the similarity as 0 for the matching vehicles outside the time window. Combining similarity matrix and filter matrix, the new developed Re-ID framework improves the matching rate to 95.7%. Furthermore, the proposed framework can even help identify vehicles that may have changed lanes, overtaken vehicles or driven on a sideroad. Such information is certainly valuable for future research on performance measure, traffic control, and congestion mitigation. Considering the significance of the trajectory data to nowadays traffic control and management and popularity of today’s surveillance cameras, this research certainly will contribute to the improvement of arterial traffic performance measure and efficient control.

In this paper, we propose a novel one-stage detector that can simultaneously detect both pedestrians and their faces. The framework is named as VikingDet for its simple but effective two-headed architecture. To tackle the challenges of person and face detection especially under surveillance cameras (e.g. low data quality, complex environments, requirements for efficiency, etc.), we make contributions in the following several aspects: 1) integrating both person and face detection into one network which current leading object detection algorithms are seldomly able to handle; 2) emphasizing detection in low-quality images. we introduce multiple thresholds for matching different sized positive samples, and set proper hyper-parameters, hence our VikingDet is able to locate small objects in surveillance cameras even of low-quality; 3) introducing a training strategy to utilize datasets on hand. Since most available public datasets annotate only people without their faces or faces without bodies, we use multi-step training and an integrated loss function to train VikingDet with these partly annotated data. As a consequence, our detector achieves satisfactory performances in several relative benchmarks with a speed at more than 60 FPS on NVIDIA TITAN X GPU, and can be further deployed on an embedded device such as NVIDIA Jetson TX1 or TX2 with a real-time speed of over 28 FPS.