Neural networks are vulnerable to adversarial perturbations: slight changes to inputs that can result in unexpected outputs. In neural network control systems, these inputs are often noisy sensor readings. In such settings, natural sensor noise – or an adversary who can manipulate them – may cause the system to fail. In this paper, we introduce the first technique to provably compute the minimum magnitude of sensor noise that can cause a neural network control system to violate a safety property from a given initial state. Our algorithm constructs a tree of possible successors with increasing noise until a specification is violated. We build on open-loop neural network verification methods to determine the least amount of noise that could change actions at each step of a closed-loop execution. We prove that this method identifies the unsafe trajectory with the least noise that leads to a safety violation. We evaluate our method on four systems: the Cart Pole and LunarLander environments from OpenAI gym, an aircraft collision avoidance system based on a neural network compression of ACAS Xu, and the SafeRL Aircraft Rejoin scenario. Our analysis produces unsafe trajectories where deviations under $1{\rm{\% }}$ of the sensor noise range make the systems behave erroneously.

A laser stripe sensor has two kinds of calibration methods. One is based on the homography model between the laser stripe plane and the image plane, which is called the one-step calibration method. The other is based on the simple triangular method, which is named as the two-step calibration method. However, the geometrical meaning of each element in the one-step calibration method is not clear as that in the two-step calibration method. A novel mathematical derivation is presented to reveal the geometrical meaning of each parameter in the one-step calibration method, and then the comparative study of the one-step calibration method and the two-step calibration method is completed and the intrinsic relationship is derived. What is more, a one-step calibration method is proposed with 7 independent parameters rather than 11 independent parameters. Experiments are conducted to verify the accuracy and robust of the proposed calibration method.

Improving designers’ ability to identify manufacturing constraints during design can help reduce the time and cost involved in the development of new products. Different design for additive manufacturing (DfAM) tools exist, but the design outcomes produced using such tools are often evaluated without comparison to existing tools. This study addresses the research gap by directly comparing design performance using two design support tools: a worksheet listing DfAM principles and a manufacturability analysis software tool that analyzes compliance with the same principles. In a randomized-controlled study, 49 nonexpert designers completed a design task to improve the manufacturability of a 3D-printed part using either the software tool or the worksheet tool. In this study, design outcome data (creativity and manufacturability) and design process data (task load and time taken) were measured. We identified statistically significant differences in the number of manufacturability violations in the software and worksheet groups and the creativity of the designs with novel build orientations. Results demonstrated limitations associated with lists of principles and highlighted the potential of software in promoting creativity by encouraging the exploration of alternative build orientations. This study provides support for using software to help designers, particularly nonexpert designers who rely on trial and error during design, evaluate the manufacturability of their designs more effectively, thereby promoting concurrent engineering design practices.

In order to ensure safe and comfortable human–robot navigation in close proximity, it is imperative for robots to possess the capability to understand human behavioral intention. With this objective in mind, this paper introduces a Human-Aware Navigation (HAN) algorithm. The HAN system combines insights from studies on human detection, social behavioral model, and behavior prediction, all while incorporating social distance considerations. This information is integrated into a layer dedicated to human behavior intention cognition, achieved through the fusion of data from laser radar and Kinect sensors, employing Gaussian functions to account for individual private space and movement trend. To cater to the mapping requirements of the HAN system, we have reduced the computational complexity associated with traditional multilayer cost map by implementing a “first-come, first-served” expansion method. Subsequently, we have enhanced the trajectory optimization equation by incorporating an improved dynamic triangle window method that integrates human behavior intention cognition, leading to the determination of an appropriate trajectory for the robot. Finally, experimental evaluations have been conducted to assess and validate the efficacy of the human behavior intention cognition and the HAN system. The results clearly demonstrate that the HAN system outperforms the traditional Dynamic Window Approach algorithm in ensuring the safety and comfort of humans in human–robot coexistence environments.

Energy consumption in buildings, both residential and commercial, accounts for approximately 40% of all energy usage in the United States, and similar numbers are being reported from countries around the world. This significant amount of energy is used to maintain a comfortable, secure, and productive environment for the occupants. So, it is crucial that energy consumption in buildings must be optimized, all the while maintaining satisfactory levels of occupant comfort, health, and safety. Machine learning (ML) has been proven to be an invaluable tool in deriving important insights from data and optimizing various systems. In this work, we review some of the most promising ways in which ML has been leveraged to make buildings smart and energy-efficient. For the convenience of readers, we provide a brief introduction to the relevant ML paradigms and the components and functioning of each smart building system we cover. Finally, we discuss the challenges faced while implementing machine learning algorithms in smart buildings and provide future avenues for research in this field.

Dual-robot system has been widely applied to the field of handling and palletizing for its high efficiency and large workspace. It is one of the key problems of the trajectory planning to determine the collision avoidance method of the dual-robot system. In the present study, a collision avoidance trajectory planning method for the dual-robot system was proposed on the basis of a modified artificial potential field (APF) algorithm. The interference and collision criterion of the dual-robot system was given firstly, which was established based on the method of kinematic analysis in robotics. And then, in consideration of the problem of excessive virtual potential field force induced by using the traditional APF algorithm in the process of dual-robot trajectory planning, a modified APF algorithm was proposed. Finally, the modified APF algorithm was used for motion control of a dual-robot palletizing process, and the collision avoidance performance of the proposed collision avoidance algorithm was studied through a dual-robot palletizing simulation and experiment. The results have shown that with the proposed collision avoidance trajectory planning algorithm, the two robots in dual-robot system can maintain a safe distance at all times during palletizing process. Compared with the traditional APF and rapidly-exploring random tree (RRT) algorithm, the trajectory solution time of the modified APF algorithm is greatly reduced. And the modified APF algorithm’s convergence time is 14.2% shorter than that of the traditional APF algorithm.

In the current investigation, a two-stage hybridization model has been used for the motion planning of humanoids in complex environmental conditions using regression analysis and the firefly algorithm. In the first step, sensory outputs are fed to the regression model, and an initial turning angle (ITA) is obtained. In the second step, the ITA is again fed as input to the firefly model along with other required inputs, and the final turning angle (FTA) is obtained. The FTA serves as the guiding parameter for the humanoids to reach their desired destinations. The developed motion planning scheme has been implemented on NAO humanoid robots in simulation and experimental platforms. A Petri-Net control strategy has been integrated along with the hybrid scheme while negotiating multiple humanoids in a common platform. The results obtained from the motion planning analysis in simulation and experimental arenas are compared against each other in terms of selected navigational parameters and observed satisfactory agreements. Finally, the proposed hybrid controller is also tested against another standard navigational model and substantial enhancement in the performance has been noticed.

Bag manipulation through robots is complex and challenging due to the deformability of the bag. Based on the dynamic manipulation strategy, we propose a new framework, ShakingBot, for the bagging tasks. ShakingBot utilizes a perception module to identify the key region of the plastic bag from arbitrary initial configurations. According to the segmentation, ShakingBot iteratively executes a novel set of actions, including Bag Adjustment, Dual-arm Shaking, and One-arm Holding, to open the bag. The dynamic action, Dual-arm Shaking, can effectively open the bag without the need to take into account the crumpled configuration. Then, the robot inserts the items and lifts the bag for transport. We perform our method on a dual-arm robot and achieve a success rate of 21/33 for inserting at least one item across various initial bag configurations. In this work, we demonstrate the performance of dynamic shaking action compared to the quasi-static manipulation in the bagging task. We also show that our method generalizes to variations despite the bag’s size, pattern, and color. Supplementary material is available at https://github.com/zhangxiaozhier/ShakingBot.

Inverse kinematics of robot is the basis of robot assembly, which directly determines the pose of robot. Because the traditional inverse solution algorithm is limited by the robot topology structure, singular pose and inverse solution accuracy, it affects the use of robots. In order to solve the above problems, an improved particle swarm optimization (PSO) algorithm is proposed to solve the inverse problem of robot. This algorithm initializes the particle population based on joint angle limitations, accelerating the convergence speed of the algorithm. In order to avoid falling into local optima and premature convergence, we have proposed a nonlinear weight strategy to update the speed and position of particles, enhancing the algorithm’s search ability, in addition introducing a penalty function to eliminate particles exceeding joint limits. Finally, the positions of common points and singular points are selected on PUMA 560 robot and redundant robot for inverse kinematics simulation verification. The results show that, compared with other algorithms, the improved PSO algorithm has higher convergence accuracy and better convergence speed in solving the inverse solution, and the algorithm has certain universality, which provides a new solution for the inverse kinematics solution of the assembly robot.

Most of the currently available three-degree-of-freedom manipulators are light load and cannot achieve full continuous rotation; given this, we designed a heavy-load manipulator that achieves unrestricted and continuous rotation. Due to manufacturing and assembly errors, parameter deviations between the real manipulator and its underlying theoretical model were unavoidable. Because of the lack of high-precision, high-frequency, and real-time closed-loop detection methods, we proposed a type of kinematics calibration of parameterized ant colony optimization and feedforward control methods. This was done to achieve high-precision motion control. First, an error model combining structural parameters and joint output angles was established, and the global sensitivity of each error source was analyzed to distinguish both primary and secondary sources. Based on the measured data of a laser tracker, the ant colony optimization was then used to identify six error sources. This resulted in both link length and joint driving errors of the designed manipulator. As it is a type of systematic error, the rounding error of the theoretical trajectory was carefully analyzed, and feedforward control methods with different coefficients were designed to further improve positioning accuracy based on the kinematic calibration. Experimental results showed that the proposed kinematic calibration and feedforward control methods achieved relatively precise motion control for the designed manipulator.

This paper presents the design, modeling, and control of a novel soft-rigid knee joint robot (SR-KR) for assisting motion. SR-KR is proposed to assist patients with knee joint injuries conducting gait training and completing walking movements. SR-KR consists of a novel soft-rigid bidirectional curl actuator, a thigh clamping structure, and a crus clamping structure. The actuating part of SR-KR is composed of soft materials, which ensures the wearing comfort and safety, while the wearing parts contain rigid structure, which ensures the efficient transmission of torque. The bending deformation model of SR-KR is established, which reveal the relationship among SR-KR’s bending curvature, working pressure, and output torque. Experiments show that SR-KR can provide more than 26.3 Nm torque for knee joint motion in human gait range. A double closed loop servo control system including attitude servo and pressure servo is built to better apply SR-KR. Mechanical property test, trajectory-driven test, and lower limb wearing test have been conducted, which show that SR-KR has ability to assist in lower limb motion and has potential in the fields of rehabilitation and human enhancement.

We design a scheme for laser-inertial odometry and mapping with bundle adjustment (BA-LIOM), which can greatly mitigate the problem of undesired ground warping due to sparsity of laser scans and significantly reduce odometry drift. Specifically, an Inertial measurement unit (IMU)-assisted adaptive voxel map initialization algorithm is proposed and elaborately integrated with the existing framework LIO-SAM, allowing for accurate registration in the beginning of the localization and mapping process. In addition, to accommodate to fast-moving and structure-less scenarios, we design a tightly coupled odometry, which jointly optimizes both the IMU preintegration constraints and scan matching with adaptive voxel maps. The voxels (edge and plane, respectively) are updated with BA optimization. And then the accurate mapping result is obtained by performing local BA. The proposed BA-LIOM is thoroughly assessed using datasets collected from multiple platforms over a variety of environments. Experimental results show the superiority of BA-LIOM over the state-of-the-art methods in robustness and precision, especially for large-scale scenarios. BA-LIOM improves the accuracy of localization by $61\%$ and $73\%$ on the buildings and lawn datasets, respectively, and has a $29\%$ accuracy improvement over LIO-SAM on the KITTI datasets. A supplementary video can be accessed at https://youtu.be/5l4ZFhTc2sw.

This article explores how the concept of remediation is part of digital memory work performed by young women on Instagram. While remediation has been used to make sense of the ways sites of memory are represented across time and through different media, mnemonic media practices and forms are remediated in digital memory work. This article draws on interviews, observations of Instagram activities, and focus group data to analyse how other media practices and forms are integrated into digital memory work on Instagram and mobilised by young women to make sense of their mnemonic use of the platform. The analysis focuses on how practices of digital memory work use direct remediation of material objects and remediation of the functionality of mnemonic media practices. It addresses how the comparisons participants make to other mnemonic media practices reveal how digital memory work involves negotiation of personal and public, private and professional, and the authentic and staged. In addition, it grapples with the way that sharing happy experiences and moments to produce a ‘highlight reel’ or ‘hall of fame’ in postfeminist digital culture has valuable and potentially harmful implications.

The article attempts to clarify what today constitutes communicative remembering. To revisit this basic mnemonic concept, our theoretical contribution starts from available approaches in social memory studies that assume a binary distinction between cultural and communicative modes of memory making. In contrast, we use concepts that treat them not as structural, historically and culturally distinct registers but as a repertoire of retrospection that hinges on the evoked temporal horizon and media usage. To further interrogate this practical articulation of memories, we direct our attention to the habitual, communicatively realised engagement with the past. We finally turn to the ways communicative remembering is done in digitally networked environments, which provide us with a pertinent mnemonic arena where rigid dichotomies of communicative memory versus cultural memory are eroded.

Welding is the most basic and widely used manufacturing process. Intelligent robotic welding is an area that has received much consideration owing to the widespread use of robots in welding operations. With the dawn of Industry 4.0, machine learning is substantially developing to alleviate issues around applying robotic welding intelligently. Identifying the correct weld joint type is essential for intelligent robotic welding. It affects the quality of the weldment and impacts the per-unit cost. The robot controller must change different welding parameters per joint type to attain the desired weld quality. This article presents an approach that uses image features like edges, corners, and blobs to identify different weld joint types using machine learning algorithms. Feature extractors perform the task of feature extraction. The feature extractor choice is crucial for accurate weld joint identification. The present study compares the performance of five feature extractors, namely (1) Histogram of gradients, (2) Local binary pattern, (3) ReLU3 layer, (4) ReLU4 layer, and (5) Pooling layer of ResNet18 Neural network applied to classifiers like Support Vector machines, K-Nearest Neighbor and Decision trees. We trained and tested the proposed model using the Kaggle Weld joint dataset (for Butt and Fillet Joints) and our in-house dataset (for Vee, lap, and corner joints). The experimental findings show that out of the 15 models, the pre-trained ResNet18 feature extractor with an Support Vector Machines classifier has excellent performance with a threefold recognition accuracy of 98.74% for the mentioned dataset with a computation time of 31 ms per image.

User models that can directly use and learn how to do tasks with unmodified interfaces would be helpful in system design to compare task knowledge and times between interfaces. Including user errors can be helpful because users will always make mistakes and generate errors. We compare three user models: an existing validated model that simulates users’ behavior in the Dismal spreadsheet in Emacs, a newly developed model that interacts with an Excel spreadsheet, and a new model that generates and fixes user errors. These models are implemented using a set of simulated eyes and hands extensions. All the models completed a 14-step task without modifying the system that participants used. These models predict that the task in Excel is approximately 20% faster than in Dismal, including suggesting why, where, and how much Excel is a better design. The Excel model predictions were compared to newly collected human data (N = 23). The model’s predictions of subtask times correlate well with the human data (r2 = .71). We also present a preliminary model of human error and correction based on user keypress errors, including 25 slips. The predictions to data comparison suggest that this interactive model that includes errors moves us closer to having a complete user model that can directly test interface design by predicting human behavior and performing the task on the same interface as users. The errors from the model’s hands also allow further exploration of error detection, error correction, and different knowledge types in user models.

This paper is concerned with the problem of collision-free path planning for manipulators in multi-obstacle scenarios. Aiming at overcoming the deficiencies of existing algorithms in excessive time consumption and poor expansion quality, a path planning algorithm named Fast Bi-directional Rapidly-exploring Random Tree (FBi-RRT) with novel heuristic node expansion is proposed, which includes a selective-expansion strategy and a vertical-exploration strategy. The selective-expansion strategy is designed to guide the selection of the nearest-neighbor node to avoid the repeated expansion failure, thereby shortening the overall planning time. Also, the vertical-exploration strategy is developed to regulate the expansion direction of the collision nodes to escape from the obstacle space with less blindness, thus improving the expansion quality and further reducing time cost. Compared with previous planning algorithms, FBi-RRT can generate a feasible path for manipulators in a drastically shorter time. To validate the effectiveness of the proposed heuristic node expansion, FBi-RRT is conducted on a 6-DOF manipulator and tested in five scenarios. The experimental results demonstrate that FBi-RRT outperforms the existing methods in time consumption and expansion quality.