Learning Speed Research Articles

Mitigating Accuracy-Robustness Trade-Off via Balanced Multi-Teacher Adversarial Distillation.

Adversarial Training is a practical approach for improving the robustness of deep neural networks against adversarial attacks. Although bringing reliable robustness, the performance towards clean examples is negatively affected after Adversarial Training, which means a trade-off exists between accuracy and robustness. Recently, some studies have tried to use knowledge distillation methods in Adversarial Training, achieving competitive performance in improving the robustness but the accuracy for clean samples is still limited. In this paper, to mitigate the accuracy-robustness trade-off, we introduce the Balanced Multi-Teacher Adversarial Robustness Distillation (B-MTARD) to guide the model's Adversarial Training process by applying a strong clean teacher and a strong robust teacher to handle the clean examples and adversarial examples, respectively. During the optimization process, to ensure that different teachers show similar knowledge scales, we design the Entropy-Based Balance algorithm to adjust the teacher's temperature and keep the teachers' information entropy consistent. Besides, to ensure that the student has a relatively consistent learning speed from multiple teachers, we propose the Normalization Loss Balance algorithm to adjust the learning weights of different types of knowledge. A series of experiments conducted on three public datasets demonstrate that B-MTARD outperforms the state-of-the-art methods against various adversarial attacks.

A robust variational mode decomposition based deep random vector functional link network for dynamic system identification

The complexity of system identification problems has been escalated due to their diverse range of applications. In this paper, the non-linear system identification problem is addressed by proposing a deep random vector functional link network (Deep-RVFLN) based on the optimized variational mode decomposition (OVMD). The proposed method has a faster learning speed and trains the network accurately without tuning parameters. Introducing a random link network connecting the input and output layers may lead to reduction in model complexity. To enhance the accuracy and reduce errors, a random vector functional link network (RVFLN) has been implemented with an increased number of hidden layers. The variational mode decomposition (VMD) algorithm is applied to decompose the signal and select optimum modes using an improved particle swarm optimization (IPSO) algorithm. In this method, the data fidelity factor (α) and the number of decomposition modes (k) are chosen by a new discrete Teaser energy operator (DTEO). The DTEO algorithm is utilized to estimate Teaser energy and it serves as a dependable indicator of overall system reliability. To test the efficacy of the model, three complex non-linear benchmark models named autoregressive (AR), moving average (MA), and autoregressive moving average (ARMA) have been considered with examples 1, 2, and 3 respectively. Based on the results and analysis, the proposed method was found to be better than other state-of-the-art methods. Finally, the proposed Deep-RVFLN identifier is implemented on a high-speed reconfigurable field-programmable gate array (FPGA) to validate the efficacy of the proposed method for non-linear system identification in the hardware platform.

Optimization of Interference Identification for Wireless Communication Based on Global Convolutional Neural Network

With the widespread adoption of smartphones and the rise of 5G technology, wireless networks fully show their bright future. In order to study the interference factors and optimization methods in the process of wireless communication, this paper studies the wireless communication interference recognition algorithm and power allocation algorithm based on global convolutional neural network, proposes a wireless communication interference recognition system model based on global convolutional neural network, and trains the system. The specific conclusions are as follows: (1) The basic modulation recognition neural network algorithm is constructed based on deep learning theory; (2) As the number of D2D users increases, the time used by the WMMSE algorithm and the equal power allocation algorithm in the traditional method increases. (3) The QL algorithm-based on greedy- is compared with the system-based QL algorithm proposed in this paper. On 200-time nodes, the SINK value of the QL algorithm based on -greedy is increased from 2.6 to 3.3, and the convergence speed is also greatly improved. The system proposed in this paper has a faster learning speed, can accelerate the learning of strategies, and can save a lot of computing resources, and effectively reduce the complexity of operations.

Online Topological Mapping on a Quadcopter with Fast Growing Neural Gas

This paper presents an online topological mapping method on a quadcopter with fast-growing neural gas. Recently, perceiving the real world in 3D space has become increasingly important, and robotics is no exception. Quadcopters are the most common type of robot working in 3D space. The ability to perceive 3D space is even required in order to enable real-time autonomous control. Dense maps are simply unpractical, while sparse maps are not suitable due to a lack of appropriate information. Topological maps then offer a balance between computational cost and accuracy. One of the most well-known unsupervised learning methods for topological mapping is growing neural gas (GNG). Unfortunately, it is difficult to increase the learning speed due to the traditional iterative method. Consequently, we propose a new method for topological mapping, called simplified multi-scale batch-learning GNG, by applying a mini-batch strategy in the learning process. The proposed method has been implemented on a quadcopter for indoor mapping applications. In addition, the topological maps are also combined with the tracking data of the quadcopter to generate a new global map. The combination is simple yet robust, based on rotation and translation strategies. Thus, the quadcopter is able to run the algorithms in real-time and maintain its performance above 30 fps.

A Fuzzy-Immune-Regulated Single-Neuron Proportional–Integral–Derivative Control System for Robust Trajectory Tracking in a Lawn-Mowing Robot

This paper presents the constitution of a computationally intelligent self-adaptive steering controller for a lawn-mowing robot to yield robust trajectory tracking and disturbance rejection behavior. The conventional fixed-gain proportional–integral–derivative (PID) control procedure lacks the flexibility to deal with the environmental indeterminacies, coupling issues, and intrinsic nonlinear dynamics associated with the aforementioned nonholonomic system. Hence, this article contributes to formulating a self-adaptive single-neuron PID control system that is driven by an extended Kalman filter (EKF) to ensure efficient learning and faster convergence speeds. The neural adaptive PID control formulation improves the controller’s design flexibility, which allows it to effectively attenuate the tracking errors and improve the system’s trajectory tracking accuracy. To supplement the controller’s robustness to exogenous disturbances, the adaptive PID control signal is modulated with an auxiliary fuzzy-immune system. The fuzzy-immune system imitates the automatic self-learning and self-tuning characteristics of the biological immune system to suppress bounded disturbances and parametric variations. The propositions above are verified by performing the tailored hardware in the loop experiments on a differentially driven lawn-mowing robot. The results of these experiments confirm the enhanced trajectory tracking precision and disturbance compensation ability of the prescribed control method.

Data-driven distributed output consensus control for multi-agent systems with unknown internal state

This paper discusses the topic of optimal output consensus control of linear time-invariant discrete-time multi-agent systems (MASs). The attainment of optimal output consensus control for MASs hinges upon the resolution of the interconnected Hamilton–Jacobi-Bellman equation, a task typically precluded by the inherent intractability of analytical solutions. Furthermore, most real-world systems are too complex to obtain the internal states of systems. To address these issues, a modified deep Q-learning network is constructed using current and historical system data rather than a precise model of the system. First, reconstructing the internal state of each agent using an adaptive distributed observer based on output feedback prevents the system instability brought on by the augmented systems. Then the local error system of the agent can be redefined. Based on the redefined error system, a data-driven adaptive dynamic programming (ADP) method is introduced, realized by using the actor–critic neural network structure. In addition, an experience replay strategy is proposed to reduce the propagation of estimation bias and improve the learning speed. Finally, the comparative numerical simulations substantiate the efficacy of the proposed algorithm in a quantifiable manner.

A novel digital intervention for improving cognitive impairment in patients with chronic schizophrenia: A randomized clinical trial

BackgroundThere is an unmet need for stand-alone digital therapeutics for cognitive impairment in schizophrenia. This study aimed to evaluate the efficacy and acceptability of a novel digital therapeutic, IBT-SC02, for cognitive impairment in stable schizophrenia patients. MethodsA randomized, parallel-group trial was conducted at the Sichuan Province Institute of Mental Health, China. Participants aged 18–50 diagnosed with schizophrenia were randomized to either the IBT-SC02 intervention or a wait-list control. The primary outcome was cognitive performance measured using the MATRICS Consensus Cognitive Battery (MCCB) composite score. ResultsA total of 80 patients were randomized (40 intervention, 40 control). The dropout rate was 5 %. The intervention group exhibited significant improvements in the MCCB composite score compared to the control, though the improvement lessened after excluding data collected by unmasked raters. Post-hoc analyses revealed that participants in the intervention group improved in four out of the seven MCCB domains (speed of processing, verbal learning, visual learning and reasoning and problem solving). No adverse events were reported. ConclusionsThese results suggest that IBT-SC02, a fully automated digital therapeutic, improved cognitive performance in patients with stable schizophrenia and can potentially be a treatment option for patients without access to trained professionals.

Transformation of Teaching through Co-Teaching and Innovative Methods

Abstract Introduction: The integration of digital technologies and virtual co-teaching has become a common practice in education. This was accelerated by the COVID-19 pandemic. These innovations play a key role in addressing global challenges and are a response to digital advances to which primary and secondary schools need to adaptively respond. Methods: The study used thematic analysis and a scoping review to investigate the effectiveness of virtual co-teaching in primary and secondary schools, with a focus on new digital technologies and artificial intelligence. Results: The results show that virtual co-teaching supported by the use of artificial intelligence is an effective alternative to traditional teaching methods that increases both the efficiency and flexibility of the learning process. Further research is necessary to show new possibilities and optimize the use of this method in formal education. Discussion: Pupils should be divided into smaller groups, respecting their individual needs, abilities, and speed of learning. Artificial Intelligence assistants should be involved to support teachers and pupils. Limitations: Experiments should be implemented in other subjects in primary and secondary schools. Conclusions: Virtual co-teaching has a strong potential to contribute to future innovations in education, especially when combined with artificial intelligence to enable more personalized learning. It extends the pedagogical space with a virtual component – a virtual teacher or subject matter expert.

Human-in-the-Loop Reinforcement Learning in Continuous-Action Space.

Human-in-the-loop for reinforcement learning (RL) is usually employed to overcome the challenge of sample inefficiency, in which the human expert provides advice for the agent when necessary. The current human-in-the-loop RL (HRL) results mainly focus on discrete action space. In this article, we propose a Q value-dependent policy (QDP)-based HRL (QDP-HRL) algorithm for continuous action space. Considering the cognitive costs of human monitoring, the human expert only selectively gives advice in the early stage of agent learning, where the agent implements human-advised action instead. The QDP framework is adapted to the twin delayed deep deterministic policy gradient algorithm (TD3) in this article for the convenience of comparison with the state-of-the-art TD3. Specifically, the human expert in the QDP-HRL considers giving advice in the case that the difference between the twin Q -networks' output exceeds the maximum difference in the current queue. Moreover, to guide the update of the critic network, the advantage loss function is developed using expert experience and agent policy, which provides the learning direction for the QDP-HRL algorithm to some extent. To verify the effectiveness of QDP-HRL, the experiments are conducted on several continuous action space tasks in the OpenAI gym environment, and the results demonstrate that QDP-HRL greatly improves learning speed and performance.

Expectation to rewards modulates learning emotional words: Evidence from a hierarchical Bayesian model.

In language acquisition, individuals learn the emotional value of words through external feedback. Previous studies have used emotional words as experimental materials to explore the cognitive mechanisms underlying emotional language processing, but have failed to recognize that languages are acquired in changing environments. To this end, this study aims to combine reinforcement learning with emotional word learning, using a probabilistic reversal learning task to explore how individuals acquire the valence of emotional words in a dynamically changing environment. Computational modeling on both behavioral and event-related potential (ERP) data revealed that individuals' expectations to rewards modulated the learning speed and temporal processing of emotional words, demonstrating a clear negative bias. Specifically, as the expected value increased, individuals responded faster and exhibited higher amplitudes for negative emotional words. These findings shed light on the neural mechanisms of emotional word learning in a volatile environment, highlighting the crucial role of expectations in this process and a preference for learning negative information.

Enhancing cooperative multi-agent reinforcement learning through the integration of R-STDP and federated learning

This paper introduces a novel approach to enhance the stability and efficiency of R-STDP in the context of federated learning. The primary objective is to stabilize the unbounded growth of R-STDP and make it more responsive to real-time changes. The methodology involves integrating R-STDP with Spiking Neural Networks and employing the norm of the neural network model for adjusting weighted aggregation in federated learning systems. The proposed method incorporates a mechanism where weights decay over time, depending on the duration since the agent last published its model. Additionally, the sampling time is dynamically adjusted based on the Euclidean norm, which measures the distance between the weight matrices of the agents and the server. The results demonstrate that the proposed event-triggered federated learning method significantly enhances learning speed and performance. At the same time, the dynamic aggregation interval efficiently reduces communication between the agents and the central server, especially after model convergence. This research presents a significant advancement in federated learning and offers a more stable, responsive, and efficient learning process.

비대면 수업에 대한 예비 과학 교사들의 인식 조사 연구

The purpose of this study is to survey the perception of pre-service science teachers regarding online classes. The subjects of this study were 76 pre-servide science teachers in the Chungcheong and Gyeongsang provinces, and an online survey was conducted. The survey instruments consisted of 13 questions in three categories: experience & class using digital tools, the learning effectiveness of online classes, and online and in-person experimental classes. Pre-service science teachers were found to lack the ability to use digital tools and the experience of class using digital tools. Pre-service science teachers perceived that in-person classes were the highest effective, but they had very positive perception of blended teaching methods. They perceived the advantages of online learning as learning at a desired location, controlling learning speed, and repeatable learning, while the disadvantages included the lack of interaction, difficulty in self-regulating learning, and poor online learning environment. In addition, they were found to have a very negative perception of online experimental classes, but showed a somewhat positive perception of learning basic concepts. In order to prepare for future school science education and enhance digital capabilities of pre-service science teachers, it is necessary to improve the curriculum for them and develop and utilize teaching-learning methods and strategies.

Learning from different perspectives for regret reduction in reinforcement learning: A free energy approach

Reinforcement learning (RL) is the core method for interactive learning in living and artificial creatures. Nevertheless, in contrast to humans and animals, artificial RL agents are very slow in learning and suffer from the curse of dimensionality. This is partially due to using RL in isolation; i.e. lack of social learning and social diversity. We introduce a free energy-based social RL for learning novel tasks. Society is formed by the learning agent and some diverse virtual ones. That diversity is in their perception while all agents use the same interaction samples for learning and share the same action set. Individual difference in perception is mostly the cause of perceptual aliasing however, it can result in virtual agents’ faster learning in early trials. Our free energy method provides a knowledge integration method for the main agent to benefit from that diversity to reduce its regret. It rests upon Thompson sampling policy and behavioral policy of main and virtual agents. Therefore, it is applicable to a variety of tasks, discrete or continuous state space, model-free, and model-based tasks as well as to different reinforcement learning methods. Through a set of experiments, we show that this general framework highly improves learning speed and is clearly superior to previous existing methods. We also provide convergence proof.

Biopsychosocial Impact of Multiple Sclerosis in Omani Patients: A Multicenter Comparative Study.

Background/Objectives: Multiple sclerosis (MS) is a chronic neurological disorder characterized by various clinical presentations and manifestations that include biopsychosocial impediments. This study has three interrelated goals relevant to biopsychosocial functioning: (i) compare reasoning ability, neuropsychological functioning, affective range, and quality of life (QoL) between people with multiple sclerosis (PwMS) and healthy controls; (ii) explore gender differences in reasoning ability and neuropsychological functioning, affective symptoms, and QoL among PwMS; and (iii) examine the relationship between QoL and cognitive performance in PwMS, focusing on those with inadequate vs. adequate QoL. Methods: This multicenter study was carried out among clinically stable PwMS (no relapse in the last two months) at follow-up in two tertiary care units in urban Oman. Healthy controls, matched for age and sex, were also recruited as a comparison group. Data were collected using cognitive batteries sensitive to current reasoning ability and conventional neuropsychological batteries designed to measure verbal learning, visual-spatial ability, and processing speed. The affective range (anxiety and depressive symptoms) and quality of life (QoL) were also evaluated. Results: The PwMS group scored lower on current reasoning ability, verbal learning, visual-spatial ability, and processing speed compared to the control group. The incidence of anxiety was higher in the PwMS group, but there were no statistically significant differences in depressive symptoms. No significant differences were found in cognitive variables between the two sexes, except in visual-spatial ability, where women outperformed men. PwMS with low QoL scored lower on attention and concentration indices than those with adequate QoL. According to QoL, no significant differences were observed in reasoning, verbal learning, or visual-spatial ability. Conclusions: The present sentinel study suggests that the Omani cohort with MS tends to have lower indices of current reasoning ability, visual and spatial memory, and cognitive speed compared to control subjects. Gender differences are minimal, except for visual-spatial abilities, where women outperform men. Quality of life significantly affects cognitive functioning. In general, the biopsychosocial impediment appears to be significant, indicating the need for comprehensive evaluation and care in the management of MS.

Invariant Extended Kalman Filtering for Pedestrian Deep-Inertial Odometry

Abstract. Indoor localization for pedestrians, which relies solely on inertial odometry, has been a topic of great interest. Its significance lies in its ability to provide positioning solutions independently, without the need for external data. Although traditional strap-down inertial navigation shows rapid drift, the introduction of pedestrian dead reckoning (PDR), and artificial intelligence (AI) has enhanced the applicability of inertial odometry for indoor localization. However, inertial odometry continues to be affected by drift, inherent to the nature of dead reckoning. This implies that even a slight error at a given moment can lead to a significant decrease in accuracy after continuous integration operations. In this paper, we propose a novel approach aimed at enhancing the positioning accuracy of inertial odometry. Firstly, we derive a learning-based forward speed using inertial measurements from a smartphone. Unlike mainstream methods where the learned speed is directly used to determine the position, we use the forward speed combined with non-holonomic constraint (NHC) as a measurement to update the state predicted within a strap-down inertial navigation framework. Secondly, we employ an invariant extended Kalman filter (IEKF)-based state estimation to facilitate fusion to cope with the nonlinearity arising from the system and measurement model. Experimental tests are carried out in different scenarios using an iPhone 12, and traditional methods, including PDR, robust neural inertial navigation (RONIN), and the EKF-based method, are compared. The results suggest that the method we propose surpasses these traditional methods in performance.

Learning‐based tracking control of AUV: Mixed policy improvement and game‐based disturbance rejection

AbstractA mixed adaptive dynamic programming (ADP) scheme based on zero‐sum game theory is developed to address optimal control problems of autonomous underwater vehicle (AUV) systems subject to disturbances and safe constraints. By combining prior dynamic knowledge and actual sampled data, the proposed approach effectively mitigates the defect caused by the inaccurate dynamic model and significantly improves the training speed of the ADP algorithm. Initially, the dataset is enriched with sufficient reference data collected based on a nominal model without considering modelling bias. Also, the control object interacts with the real environment and continuously gathers adequate sampled data in the dataset. To comprehensively leverage the advantages of model‐based and model‐free methods during training, an adaptive tuning factor is introduced based on the dataset that possesses model‐referenced information and conforms to the distribution of the real‐world environment, which balances the influence of model‐based control law and data‐driven policy gradient on the direction of policy improvement. As a result, the proposed approach accelerates the learning speed compared to data‐driven methods, concurrently also enhancing the tracking performance in comparison to model‐based control methods. Moreover, the optimal control problem under disturbances is formulated as a zero‐sum game, and the actor‐critic‐disturbance framework is introduced to approximate the optimal control input, cost function, and disturbance policy, respectively. Furthermore, the convergence property of the proposed algorithm based on the value iteration method is analysed. Finally, an example of AUV path following based on the improved line‐of‐sight guidance is presented to demonstrate the effectiveness of the proposed method.

Student perspectives on the flipped classroom concept in secondary math lessons

The flipped classroom concept is the reversal of the traditional teaching activities so that new content is first worked through by the learners independently and then the joint lesson time is used to practise and consolidate this content. This concept can result in didactic advantages. However, it is assumed that the concept also places increased demands on learners. As part of an intervention study with 452 participants in 15 lower secondary schools in Lower Saxony (Germany), pupils were given a questionnaire including the following items “I like this in the flipped classroom” and “I don’t like this in the flipped classroom”. This article analyses the answers to those questions using a qualitative content analysis. The resulting categories are analysed in more detail in a thematic analysis to identify specific aspects of the flipped classroom that the learners like or dislike. The preparation through videos, self-directed learning and adaptability of the learning speed were named particularly frequently in the positively connoted categories. The negative aspects included comprehension problems, the desire for more explanations from the teachers, time pressure, and time management problems. These results contrast with previous research results on the flipped classroom concept in higher education didactics.

Macro- and Microstructural White Matter Differences in Neurologic Postacute Sequelae of SARS-CoV-2 Infection.

Neuropsychiatric complications of SARS-CoV-2 infection, also known as neurologic postacute sequelae of SARS-CoV-2 infection (NeuroPASC), affect 10%-60% of infected individuals. There is growing evidence that NeuroPASC is a multi system immune dysregulation disease affecting the brain. The behavioral manifestations of NeuroPASC, such as impaired processing speed, executive function, memory retrieval, and sustained attention, suggest widespread WM involvement. Although previous work has documented WM damage following acute SARS-CoV-2 infection, its involvement in NeuroPASC is less clear. We hypothesized that macrostructural and microstructural WM differences in NeuroPASC participants would accompany cognitive and immune system differences. In a cross-sectional study, we screened a total of 159 potential participants and enrolled 72 participants, with 41 asymptomatic controls (NoCOVID) and 31 NeuroPASC participants matched for age, sex, and education. Exclusion criteria included neurologic disorders unrelated to SARS-CoV-2 infection. Assessments included clinical symptom questionnaires, psychometric tests, brain MRI measures, and peripheral cytokine levels. Statistical modeling included separate multivariable regression analyses of GM/WM/CSF volume, WM microstructure, cognitive, and cytokine concentration between-group differences. NeuroPASC participants had larger cerebral WM volume than NoCOVID controls (β = 0.229; 95% CI: 0.017-0.441; t = 2.16; P = .035). The most pronounced effects were in the prefrontal and anterior temporal WM. NeuroPASC participants also exhibited higher WM mean kurtosis, consistent with ongoing neuroinflammation. NeuroPASC participants had more self-reported symptoms, including headache, and lower performance on measures of attention, concentration, verbal learning, and processing speed. A multivariate profile analysis of the cytokine panel showed different group cytokine profiles (Wald-type-statistic = 44.6, P = .046), with interferon (IFN)-λ1 and IFN-λ2/3 levels higher in the NeuroPASC group. NeuroPASC participants reported symptoms of lower concentration, higher fatigue, and impaired cognition compatible with WM syndrome. Psychometric testing confirmed these findings. NeuroPASC participants exhibited larger cerebral WM volume and higher WM mean kurtosis than NoCOVID controls. These findings suggest that immune dysregulation could influence WM properties to produce WM volume increases and consequent cognitive effects and headaches. Further work will be needed to establish mechanistic links among these variables.

Enhancing learning experiences: EEG-based passive BCI system adapts learning speed to cognitive load in real-time, with motivation as catalyst.

Computer-based learning has gained popularity in recent years, providing learners greater flexibility and freedom. However, these learning environments do not consider the learner's mental state in real-time, resulting in less optimized learning experiences. This research aimed to explore the effect on the learning experience of a novel EEG-based Brain-Computer Interface (BCI) that adjusts the speed of information presentation in real-time during a learning task according to the learner's cognitive load. We also explored how motivation moderated these effects. In accordance with three experimental groups (non-adaptive, adaptive, and adaptive with motivation), participants performed a calibration task (n-back), followed by a memory-based learning task concerning astrological constellations. Learning gains were assessed based on performance on the learning task. Self-perceived mental workload, cognitive absorption and satisfaction were assessed using a post-test questionnaire. Between-group analyses using Mann-Whitney tests suggested that combining BCI and motivational factors led to more significant learning gains and an improved learning experience. No significant difference existed between the BCI without motivational factor and regular non-adaptive interface for overall learning gains, self-perceived mental workload, and cognitive absorption. However, participants who undertook the experiment with an imposed learning pace reported higher overall satisfaction with their learning experience and a higher level of temporal stress. Our findings suggest BCI's potential applicability and feasibility in improving memorization-based learning experiences. Further work should seek to optimize the BCI adaptive index and explore generalizability to other learning contexts.

Reinforcement Learning with Decoupled State Representation for Robot Manipulations

Deep reinforcement learning has significantly advanced robot manipulations by providing an alternative solution for designing control strategies using raw images as direct inputs. While images offer additional environmental information, the end-to-end policy training manner (from image to action) requires simultaneous representation and task learning by the agent. This often necessitates a substantial number of interaction samples to achieve satisfactory policy performance. Previous works has attempted to address this challenge by learning a visual representation model that encodes the entire image into a low-dimensional vector before the policy training. However, since this vector contains both robot and object information, it inevitably introduces coupling within the state, which can mislead the policy training process. In this study, a novel method called Reinforcement Learning with Decoupled State Representation is proposed to effectively decouple robot and object information within the state representation. Experimental results demonstrate that the proposed method exhibits faster learning speed and achieves superior performance compared to previous methods across various robot manipulation tasks. Moreover, with only 3096 offline images, the proposed method successfully applies to real-world robot pushing tasks, which demonstrates its high practicability.