A light-weight, real-time approach named RTGN (real-time grasp net) was proposed to improve the accuracy and speed of robotic grasp detection for novel objects of diverse shapes, types and sizes. Firstly, a multi-scale dilated convolution module was designed to construct a light-weight feature extraction backbone. Secondly, a mixed attention module was designed to help the network focus more on meaningful features. Finally, the pyramid pool module was deployed to fuse the multi-level features extracted by the network, thereby improving the capability of grasp perception to the object. On the Cornell grasping dataset, RTGN generated grasps at a speed of 142 frame per second and attained accuracy rates of 98.26% and 97.65% on image-wise and object-wise splits, respectively. In real-world robotic grasping experiments, RTGN obtained a success rate of 96.0% in 400 grasping attempts across 20 novel objects. Experimental results demonstrate that RTGN outperforms existing methods in both detection accuracy and detection speed. Furthermore, RTGN shows strong adaptability to variations in the position and pose of grasped objects, effectively generalizing to novel objects of diverse shapes, types and sizes.

The dense small target detection algorithm LSA_YOLO based on YOLOv5s for UAVs with complex backgrounds and multiples of small targets with dense distribution was proposed for UAV images. A multi-scale feature extraction module LM-fem was constructed to enhance the feature extraction capability of the network. A new hybrid domain attention module S-ECA relying on multi-scale contextual information has been put forward and a algorithm focus on target information was established aiming to suppress the interference of complex backgrounds. The adaptive weight dynamic fusion structure AFF was designed to assign reasonable fusion weights to both shallow and deep features. The capability of algorithm in detecting dense small targets in complex backgrounds was improved given the application of S-ECA and AFF in the structure of PANet. The loss function Focal-EIOU was utilized instead of the loss function CIOU to accelerate model detection efficiency. Experimental results on the public dataset VisDrone2021 public dataset show that the average detection accuracy for all target classes improves from 51.5% for YOLOv5s to 57.6% for LSA_YOLO when the set input resolution is set to 1 504 × 1 504.

Aiming at the problems of data sparsity, cold start, low interpretability of recommendation, and insufficient personalization in recommender system, the integration of knowledge graph into recommender system was analyzed. From the demand of recommender system, the concept of knowledge graph, and the integration approach of recommender system and knowledge graph, the problems of current recommender system and the solutions of recommender system after integrating knowledge graph were summarized. It was reviewed that, in recent years, the attention mechanism, neural network and reinforcement learning methods were combined, by which the principles of node trade-off, node integration, and paths exploring were used to make full use of the complex structural information in knowledge graph, so as to improve the satisfaction degree with the recommender system. The challenges and possible future development direction of the recommender system integrating the knowledge graph were put forward in terms of knowledge graph completeness, dynamics, availability of higher-order relationships, and the performance of the recommendation.

A vehicle motion planning algorithm based on deep reinforcement learning was proposed to satisfy the efficiency and comfort requirements of intelligent connected vehicles at unsignalized intersections. Temporal convolutional network (TCN) and Transformer algorithms were combined to construct the intention prediction model for surrounding vehicles. The multi-layer convolution and self-attention mechanisms were used to improve the capability of capturing vehicle motion feature. The twin delayed deep deterministic policy gradient (TD3) reinforcement learning algorithm was employed to build the vehicle motion planning model. Taking the driving intention of surrounding vehicle, driving style, interaction risk, and the comfort of ego vehicle into consideration comprehensively, the state space and reward functions were designed to enhance understanding the dynamic environment. Delaying the policy updates and smoothing the target policies were conducted to improve the stability of the proposed algorithm, and the desired acceleration was output in real-time. Experimental results demonstrated that the proposed motion planning algorithm can perceive the real-time potential interaction risk based on the driving intention of surrounding vehicles. The generated motion planning strategy met the requirements of the efficiency, safety and comfort. It showed excellent adaptability to different styles of surrounding vehicles and dense interaction scenarios, and the success rates exceeded 92.1% in various scenarios.

Inspired by the idea that feature information between different pixel-level visual tasks can guide and optimize each other, a traffic scene perception algorithm based on multi-task learning theory was proposed for joint semantic segmentation and depth estimation. A bidirectional cross-task attention mechanism was proposed to achieve explicit modeling of global correlation between tasks, guiding the network to fully explore and utilize complementary pattern information between tasks. A multi-task Transformer was constructed to enhance the spatial global representation of specific task features, implicitly model the cross-task global context relationship, and promote the fusion of complementary pattern information between tasks. An encoder-decoder fusion upsampling module was designed to effectively fuse the spatial details contained in the encoder to generate fine-grained high-resolution specific task features. The experimental results on the Cityscapes dataset showed that the mean IoU of semantic segmentation of the proposed algorithm reached 79.2%, the root mean square error of depth estimation was 4.485, and the mean relative error of distance estimation for five typical traffic participants was 6.1%. Compared with the mainstream algorithms, the proposed algorithm can achieve better comprehensive performance with lower computational complexity.

In order to realize gesture recognition and hand state recognition at the same time, a single inertial measurement unit-based gesture recognition and touch recognition prototype was built, considering the inertial measurement unit at high sample rate has the capability of collecting motion signals and vibration signals simultaneously. The differences within hand state data and gesture data in the time and frequency domains were visually analyzed. Hand state, slipping gesture and circling gesture data sets were established. Considering the difference within data features, differential feature extraction methods were proposed, and neural network structures for hand state classification and gesture classification were constructed. Neural network models were trained by the data sets to achieve 99% accuracy rate in the comprehensive hand state recognition task, and 98% accuracy rate in both the slipping gesture recognition task and the circling gesture recognition task. A prototype program framework for real-time data stream processing, state shifting, and unknown class judgment was proposed. And a real-time program based on the hand state recognition model entities and the gesture recognition model entities was built, and the overall computational latency of the actual operation and the single model computational latency were measured, in order to prove the capability of real-time computing. Experimental results of model evaluation and real-time computing verification showed that, accurate and real-time hand states and gesture recognition with high sample rate inertial measurement units was feasible.

In response to the need for ankle rehabilitation training, a lightweight, easy-to-wear flexible ankle exoskeleton robot was designed using modular drive units and Bowden cables through analysis of ankle joint mechanics. The robot can provide assistance for ankle plantarflexion/dorsiflexion and inversion/eversion movements. Position control and torque control are used for flexible exoskeleton during the dorsiflexion and plantarflexion stages, respectively. Position control is mainly based on traditional proportional integral derivative(PID), while torque control uses force as a feedback signal to establish an admittance model between the interaction force difference and the Bowden cable core displacement compensation. The admittance parameters are dynamically adjusted through the Sigmoid deformation function to meet the requirements of assistive torque output and human-machine interaction compliance. Experimental data showed that the position tracking error was stable within 0.46 cm, and the force output error was stable within ?1.5-1.5 N, meeting the needs of human rehabilitation training.

A behavioral decision-making model for autonomous vehicles based on driver behavior prediction was proposed, in order to solve the problem of right-of-way assignment at an unsignalized intersection in the human-machine hybrid driving environment. Fuzzy logic method was used to construct the driver's risk perception model. Then the human-driven vehicle’s behavior selection strategy was predicted based on the risk equilibrium theory and the acceptable risk interval. Finally the comprehensive utility function of autonomous vehicle was established, using game theory to solve the optimal behavior strategy combination, and realize the coordinated control of vehicles at an unsignalized intersection. Simulation results show that autonomous vehicles can effectively avoid collision accidents and improve the efficiency of autonomous vehicles passing through unsignalized intersections when facing heterogeneous drivers and guarantees that driver risk perception values are within acceptable ranges. Among the 15 experiments conducted, 93.3% of the experimental groups were able to ensure that the time difference of vehicles passing through the conflict point was larger than the acceptable safe transit interval. The travel time of the autonomous vehicle was 1.07~2.43 times that of free flow under different situations. Experimental comparison between the proposed method and the unpredicted selfish game method shows that the proposed method can significantly improve the autonomous vehicle’s traffic efficiency.

A multi-behavior aware service recommendation method based on hypergraph graph convolutional neural network (MBSRHGNN) was proposed to resolve the problem of insufficient high-order service feature extraction in existing service recommendation methods. A multi-hypergraph was constructed according to user-service interaction types and service mashups. A dual-channel hypergraph convolutional network was designed based on the spectral decomposition theory with functional and structural properties of multi-hypergraph. Chebyshev polynomial was used to approximate hypergraph convolution kernel to reduce computational complexity. Self-attention mechanism and multi-behavior recommendation methods were combined to measure the importance difference between multi-behavior interactions during the hypergraph convolution process. A hypergraph pooling method named HG-DiffPool was proposed to reduce the feature dimensionality. The probability distribution for recommending different services was learned by integrating service embedding vector and hypergraph signals. Real service data was obtained by the crawler and used to construct datasets with different sparsity for experiments. Experimental results showed that the MBSRHGNN method could adapt to recommendation scenario with highly sparse data, and was superior to the existing baseline methods in accuracy and relevance.

A path planning algorithm based on the fusion of the improved A* algorithm and the random obstacle avoidance dynamic window method (ROA-DWA) was proposed in order to address the issues of excessive traversal nodes, redundant points, non-smooth paths, lack of global guidance, susceptibility to local optima, and low safety in traditional A* algorithm and dynamic window approach (DWA) for robot path planning. The search efficiency was improved by adjusting the weights of heuristic functions, Floyd’s algorithm, redundant point deletion strategy, static and dynamic obstacle classification, and speed adaptive factor. The length of the path and the number of inflection points were reduced, and the influence of known obstacles on the path was minimized to improve the efficiency of dynamic obstacle avoidance, which enabled the robot to smoothly arrive at the target point and improved the safety of the robot, and better adapted to complex dynamic and static environments. The experimental results show that the algorithm has better global optimality and local obstacle avoidance ability, and shows better advantages in large maps.

Foundational models in natural language processing, computer vision and multimodal learning have achieved significant breakthroughs in recent years, showcasing the potential of general artificial intelligence. However, these models still fall short of human or animal intelligence in areas such as causal reasoning and understanding physical commonsense. This is because these models primarily rely on vast amounts of data and computational power, lacking direct interaction with and experiential learning from the real world. Many researchers are beginning to question whether merely scaling up model size is sufficient to address these fundamental issues. This has led the academic community to reevaluate the nature of intelligence, suggesting that intelligence arises not just from enhanced computational capabilities but from interactions with the environment. Embodied intelligence is gaining attention as it emphasizes that intelligent agents learn and adapt through direct interactions with the physical world, exhibiting characteristics closer to biological intelligence. A comprehensive survey of embodied artificial intelligence was provided in the context of foundational models. The underlying technical ideas, benchmarks, and applications of current embodied agents were discussed. A forward-looking analysis of future trends and challenges in embodied AI was offered.

An interactive visualization generation method for time series data based on transfer learning was proposed in order to address the inconsistency in data distribution across time-series data and facilitate the application of pattern analysis to other data. Transfer component analysis was applied to transfer features extracted from each time series data. The user’s analysis on one of the time series data served as labels. The classifier was trained on the source domain and applied to multiple target domains in order to achieve pattern recommendations. Two case studies and expert interviews with real-world weather data and bearing signal data were conducted to verify the effectiveness and practicality of the method by improving the efficiency of temporal data exploration and reducing the impact of inconsistent data distribution.

A new estimation network was proposed for improving the insufficient occlusion handling ability of existing human pose estimation methods. An occluded parts enhanced convolutional network (OCNN) and an occluded features compensation graph convolutional network (OGCN) were included in the proposed network. A high-low order feature matching attention was designed to strengthen the occlusion area features, and high-adaptation weights were extracted by OCNN, achieving enhanced detection of the occluded parts with a small amount of occlusion data. OGCN strengthened the shared and private attribute compensation node features by eliminating the obstacle features. The adjacency matrix was importance-weighted to enhance the quality of the occlusion area features and to improve the detection accuracy. The proposed network achieved detection accuracy of 78.5%, 67.1%, and 77.8% in the datasets COCO2017, COCO-Wholebody, and CrowdPose, respectively, outperforming the comparative algorithms. The proposed network saved 75% of the training data usage in the self-built occlusion dataset.

A driver-automation shared control strategy based on non-cooperative game (NCG) theory was proposed in order to reduce the conflict operations between the driver and intelligent system during the co-driving. The lane-keeping shared control problem was mathematically described by the first-order differential equation based on the linear two degree-of-freedom vehicle model. The NCG theory was employed to resolve the weight allocation problem of the shared control system, where the decision makers would act on the same dynamic system. The driving control authority was designed. Then the smooth transition of driving control authority between the driver and intelligent system was achieved by utilizing the preview offset distance (POD) to update the confidence matrix. The desired front wheel angle of lane-keeping shared control was transformed into an online quadratic programming problem formulated as a quadratic cost function with linear inequality constraints based on the model predictive control (MPC) framework. The shared control strategy was validated on the driver-in-the-loop CarSim/Simulink platform. Results demonstrate that such strategy can well-guarantee lateral tracking accuracy and the priority of the driver’s control authority.

A small target detection method for unmanned aerial vehicle (UAV) based on adaptive up-sampling and spatial correlation enhancement was proposed, to resolve the problem of false detection and missed detection caused by the small size of UAV and the difficulty of feature extraction under complex backgrounds. Firstly, the important contextual information was obtained by multi-scale dilated convolution, and then the attention feature fusion module was used to suppress the information conflict of multi-scale feature fusion; Secondly, a new up-sampling method of sub-pixel convolution and bilinear interpolation adaptive fusion was adopted to balance the computation and to fuse more UAV feature information; Finally, spatial correlation enhancement strategies for local and global spatial features were performed on deep features to improve the sensitivity of foreground targets in complex backgrounds and enhance target expression to suppress background noise. Ablation experiments and comparative experiments were implemented on the self-made UAV dataset. The mAP0.5 and mAP0.5:0.95 of the proposed algorithm were increased by 2.4% and 2.7% respectively, compared with those of the original YOLOv5 algorithm. Furthermore, the detection speed was able to achieve 58.5 frames per second. The performance of the proposed algorithm was also verified on the VisDrone2019 dataset, and its mAP0.5 and mAP0.5:0.95 were respectively higher than those of the YOLOv5 algorithm by 4.6% and 1.3%.

Session-based recommendation algorithms only capture users’ short-term dynamic interests, ignoring the impact of long-term interests and social friends on their behavior. To address the problem, a recommendation algorithm combining social influence and long short-term preferences was proposed. Firstly, a novel heterogeneous relation graph was designed to organize users’ social relations and historical interaction behaviors. And a heterogeneous graph neural network based on the attention mechanism was proposed to learn the graph, and to obtain long-term preference for integrating social influence of users. Moreover, considering the problem of noise caused by inconsistent social influence, a weighted and pruning strategy was proposed to reduce noise interference and enrich the graph structure information. Then, a lossless session modeling method was used to capture users’ short-term preference. Finally, users’ short-term preference and long-term preference were adaptively fused to obtain a feature representation that reflects users’ global preferences. Experimental results on Gowalla and Delicious datasets show that the indicators of the proposed method are significantly improved compared with the existing advanced methods, which proves the effectiveness of the proposed algorithm.

A new design scheme of crab-like hexapod origami robot was proposed by combining the origami structure with the multi-legged robot design and coupling Miura origami and six-fold origami aiming at the problems that the existing origami robots have a single structure and insufficient flexibility in movement. The motion configuration of the origami robot was expanded, and the motion flexibility of the origami robot was improved. Each leg of the robot has two degrees of freedom under the symmetry hypothesis. The vertices of the robot legs were treated as joints, and the crease lines were regarded as links. A planar link equivalent model of the robot legs was established with the folding angle as the motion variable. The theoretical range of motion for the robot’s foot was determined through simulation calculations. Then tapered panel technique was utilized to thicken the folding surfaces and prevent physical interference between adjacent folding surfaces. A three-dimensional model of the origami crab-like hexapod robot was constructed. The relationship between the folding angle and foot motion was analyzed based on the equivalent model of planar links, and the foot motion trajectory and gait of the robot were designed. The experimental prototype of origami bionic hexapod robot was designed and manufactured by using 3D printing technology, and the lateral movement of the robot was realized based on STM32 microcontroller control. Results show that the origami bio-inspired robot can realize the conversion from plane configuration to a crab-like configuration. The robot can move smoothly left and right under the coordinated movement of six legs.

A multi-modal augmented model for click through rate (MMa4CTR) tailored for micro-videos recommendation was proposed. Multi-modal data derived from user interactions with micro-videos were effectively leveraged to construct embedded user representations and capture diverse user interests across multi-modal. The aim was to reveal the latent semantic commonalities, by combining and crossing features across modalities. The overall recommendation performance was boosted via two training strategies, automatic learning rate adjustment and validation interruption. A computationally efficient multi-layer perceptron architecture was employed, in order to address the computational demands brought on by the vast amount of multi-modal data. Performance comparison experiments and sensitivity analyses of hyperparameter on WeChat Video Channel and TikTok datasets demonstrated that MMa4CTR outperformed baseline models, delivering superior recommendation results with minimal computational resources. Additionally, ablation studies performed on both datasets further validated the significance and efficacy of the micro-video modality cross module, the user multi-modal embedding layer, and the strategies for automatic learning rate adjustment and validation interruption in enhancing recommendation performance.

An enhanced genetic algorithm was proposed to address the challenge of area coverage path planning for a tilt-rotor unmanned aerial vehicle (TRUAV) amidst multiple obstacles. A preliminary coverage path plan for the designated task area was devised, utilizing the minimum spanning and back-and-forth path generation algorithms. The area coverage dilemma was transformed into a traveling salesman problem to optimize the sequence of the coverage path. A fishtail-shaped obstacle avoidance strategy was proposed to circumvent obstacles within the region. The nearest neighbor algorithm was introduced to generate a superior initial population than a genetic algorithm. A three-point crossover operator and a dynamic interval mutation operator were adopted in the genetic processes to improve the proposed algorithm's global search capacity and prevent the algorithm from falling into local optima. The efficacy of the proposed algorithm was rigorously tested through simulations in polygonal areas with multiple obstacles. Results showed that, compared to the sequential path coverage algorithm and the genetic algorithm, the proposed algorithm reduced the length of the coverage path by 7.80%, significantly enhancing the coverage efficiency of TRUAV in the given task areas.

A cross-domain recommendation model that utilizes source domain data augmentation and multi-interest refinement transfer was proposed in order to address the issues of difficulty in modeling interest preferences in cross-domain recommendation tasks caused by the lack of user interaction data in the source domain, as well as the problem of ignored associations between multiple interests. A source-domain data augmentation strategy was introduced, generating a denoised auxiliary sequence for each user in the source domain. Then the sparsity of user interaction data in the source domain was alleviated, and enriched user interest preferences were obtained. The interest extraction and multi-interest refinement transfer were implemented by utilizing the dual sequence multi-interest extraction module and the multi-interest refinement transfer module. Three publicly cross-domain recommendation evaluation tasks were conducted. The proposed model achieved the best performance compared with the best baseline, reducing the average MAE by 22.86% and the average RMSE by 19.65%, which verified the effectiveness of the method.