Manipulator-based machine tending is a key automation technology in computer numerical control (CNC) machining, and mobile-base configurations are increasingly adopted to enable flexible reconfiguration in high-mix, low-volume manufacturing environments. However, once the robot base is relocated, the previously taught trajectory is no longer valid, and manual reteaching is required, which is time-consuming, labor-intensive, and highly dependent on operator skill. This paper proposes a marker-based machine tending method that combines multi-marker Perspective-n-Point (PnP) pose estimation with Iterative Closest Point (ICP) registration in a two-step calibration. In the first step, PnP is used to move the robot to the pre-stored reference pose. In the second step, ICP is applied to the point cloud data obtained at the current robot pose and the pre-stored point cloud data to estimate the change in the robot base pose with respect to the CNC machine. Based on this estimate, the robot can resume its task after base relocation. Validation in a CNC lathe machine tending environment showed that, compared with conventional marker-only calibration, the position repeatability measured by a laser tracker improved from 2.658 to 1.795 mm and the machine tending success rate increased from 47% to 93%, confirming the effectiveness of the proposed method.
Large-scale surface-following tasks require mobile manipulators when the work surface exceeds the workspace of a fixedbase manipulator. Base placement must consider not only inverse kinematic reachability but also collision feasibility and joint-space continuity along the task-point sequence. This paper proposes a grid-based base placement framework for these tasks. Task points and target poses are generated from a stereolithography (STL) mesh. For each predefined base candidate, inverse kinematics and collision checks are performed, and feasible task points and their corresponding joint solutions are stored for base assignment. Joint-space discontinuity is evaluated from joint variations between adjacent task points and incorporated into the base selection cost together with the number of base changes and planar travel distance. MATLAB simulations were conducted on a vehicle roof model. Of 885 task points, 853 had collision-free inverse kinematic solutions across the predefined set of 35 base candidates, and the proposed method assigned all 853 feasible points. Compared with a collision-aware baseline method, the proposed method reduced the number of joint discontinuities from 1 to 0, the maximum joint variation from 3.071 to 0.482 rad, the number of base changes from 22 to 9, and the planar base travel distance from 13.60 to 3.54 m.
Vertical farms provide structured, controlled environments for stable crop production and suit robotic automation. However, autonomous harvesting there remains challenging because visually similar fruits are densely arranged and can occlude the target fruit. We propose a Vision-Language-Action (VLA) pipeline augmented with explicit target-object information for dualarm robotic strawberry harvesting. A perception module selects the target strawberry from wrist-camera images; its 3D position is concatenated with the robot state, and its segmentation mask is overlaid on the RGB image. The pipeline, built on GR00T N1.7, was evaluated in 220 real-robot trials under four input configurations: baseline, mask only, position only, and both. In single-object trials (45 each), success rates were 75.6%, 91.1%, 88.9%, and 95.6%, respectively; only the last differed significantly from the baseline, and the augmented configurations did not differ significantly from one another. The improvement arose mainly after failed grasps: the baseline tended to regrasp at the same location, whereas target-informed policies re-approached the target. In a three-strawberry cluster with one fruit occluded in the head-camera view, policies given the 3D position succeeded in all 5 trials, versus 2 for the baseline. These results suggest that explicit target-object information improves VLA-based harvesting under occlusion within the tested layouts.
Robotic grasping in unstructured environments demands closed-loop feedback that goes beyond conventional fixed open-loop force control. Vision sensors suffer from occlusion at the instant of contact, whereas wrist-mounted force/ torque sensors cannot resolve local shear stress distributions or incipient slip at multi-contact interfaces. Tactile sensors bridge this gap by mapping high-dimensional contact physics directly and in real time. This paper surveys grasp intelligence technology within a Perception–State Estimation–Action framework. Specifically, it reviews optical and arraytype tactile sensors; slip detection algorithms, ranging from physics-based local deformation analysis to deep spatiotemporal networks; and active closed-loop force regulation, including diffusion-policy-based reactive control. Applications in precision semiconductor and electronics assembly are then discussed. Finally, outstanding challenges in mechanical durability, ultra-low-latency data processing, intelligent tactile skin, and multi-modal Vision-Language-Action integration are identified as priorities for future research.
Although 3D shape completion facilitates robotic grasping from partial observations, deterministic completion models overlook regional uncertainty, which can result in suboptimal grasp poses. To address this limitation, we propose an uncertainty-aware robotic grasping framework that couples geometric completion confidence with physical constraints. The framework integrates a Multi-rate Dropout Ensemble (MDE) into the shape completion network, enabling it to restore 3D geometry while simultaneously quantifying epistemic uncertainty. The quantified uncertainty is then mapped onto a 3D gripper bounding box to evaluate the local reliability of each grasp candidate. Evaluations in simulation on a diverse set of unseen objects demonstrate that our approach effectively suppresses unreliable grasp hypotheses. Overall, the framework significantly reduces geometric reconstruction errors and improves grasping success rates, confirming its strong generalization capability and operational safety under occlusion.
This paper proposes a velocity-proportional virtual force-based admittance control method to suppress desired position accumulation during contact in robot direct teaching. In conventional admittance control, the force that an operator applies to the handle is converted into a torque component about the robot drive axis and then used to generate the desired position. When the end fixture is constrained by contact, however, the desired position can keep moving away from the actual position, thereby increasing the tracking error and motor torque. The proposed method detects this accumulation state on the basis of the motor torque magnitude and the sign consistency between the motor torque and the converted torque component. Upon detection, the converted torque component is replaced with a virtual component of opposite sign that is proportional to the desired velocity, thereby damping the desired velocity and limiting further desired position updates. In contact experiments on the third rotational axis of an ARF module, the conventional method reached 300% of the rated torque and caused a servo driver overload error. In contrast, the proposed method limited the maximum position error to approximately 0.086° and the maximum motor torque to approximately 63% of the rated torque, without an overload error.
This study proposes a model predictive path integral (MPPI) control method that combines adaptive covariance adjustment with low-pass perturbation filtering for the pose control of a robotic manipulator. The proposed method updates the sampling covariance according to the position and orientation errors of the tool center point, which allows a large sampling range during the approach phase and reduces the perturbation magnitude near the target pose. The sampled perturbation sequences are then passed through a low-pass filter before rollout evaluation, and the nominal control sequence is updated using the weighted sum of the filtered perturbations. This structure adaptively regulates the sampling range while reducing abrupt temporal variations in the sampled perturbation sequences. Numerical simulations were conducted on a 6-DoF robotic manipulator under a sequential target-pose scenario with five random seeds. The proposed MPPI method reached the final target-hold condition at 4.10 ± 0.08 s and maintained root-mean-square (RMS) position and orientation errors of 0.47 ± 0.12 mm and 0.05 ± 0.01°, respectively, with an RMS input variation of 0.02 ± 0.001 rad/s. These results indicate that the proposed method can maintain fast target-approach performance while improving near-goal stability and input smoothness in robotic manipulator pose control.
This paper presents a control method for a cable-suspended aerial manipulator performing high-altitude contact tasks. The target tasks include polishing and wiping on planar and curved surfaces, in which the end-effector must maintain a desired contact force while the suspended aerial manipulator remains stable against contact-induced disturbances. The proposed framework consists of a winch-based attitude and altitude controller, a propulsion controller for horizontal position and yaw stabilization, and a task-space manipulator controller with admittance-based force regulation. The winch system regulates the roll, pitch, and altitude of the platform by adjusting the cable lengths. The propulsion system counteracts horizontal disturbances through damping control, position regulation, disturbance compensation, and constrained control allocation. The manipulator generates a reference end-effector trajectory from the contact-force error along the surface-normal direction and tracks this reference using closed-loop inverse kinematics. Simulation results for side-plane, bottom-plane, side-curved, and bottom-curved contact tasks demonstrate that the proposed method maintains the desired contact force while suppressing platform displacement during contact.
Manufacturing environments implementing AI-based predictive maintenance face critical constraints imposed by limited data and label availability. This study examines two contrasting industrial cases to determine how anomaly detection model selection should align with data conditions. Case A (Direct Loom, Unsupervised): From 636,540 rows of programmable logic controller (PLC) timeseries data collected over nine days without anomaly labels, a six-stage preprocessing pipeline and sliding-window feature extraction (W = 64, stride = 16) produced 160-dimensional vectors, which were fed into an AutoEncoder (160-80-40-80-160). Five-fold time-series cross-validation yielded an anomaly detection rate (ADR) of 1.0% (±0.0%) with a mean false positive rate (FPR) of 0.2%. This FPR was well below the theoretical upper bound of 1.0% implied by the p99 threshold definition, indicating that the model discriminates beyond random chance. The inference time of 0.43 s remained well below the PLC sampling interval of 1.14 s. Case B (Press Induction Motor, Supervised): An XGBoost classifier with class-imbalance compensation, trained on 717 vibration feature records collected over 61 days with anomaly labels, achieved an F1-score of 0.870 and a ROC-AUC of 0.847. These findings support a practical stepwise strategy: begin with an unsupervised AutoEncoder to log anomalies, and then progressively transition to supervised models as labeled data accumulates.
Emotional engagement is a critical factor in audiovisual entertainment environments such as movies, performances, and immersive exhibitions. This study developed a voice-coil-based seat-integrated vibrotactile system and investigated its effects on users’ affective experiences during audiovisual content consumption. Unlike conventional wearable haptic devices, the proposed system provides vibrotactile feedback through the seat structure without requiring equipment, offering a practical configuration for public audience environments. A total of 127 participants were randomly assigned to either an audiovisualonly condition (AV) or an audiovisual condition with synchronized seat vibrotactile feedback (AVH). Participants viewed 50 emotion-inducing video clips categorized into neutral, fun, sadness, anger, and fear. Subjective responses were collected using self-assessment questionnaires measuring valence, arousal, dominance, liking, and emotion intensity. The data were analyzed using linear mixed-effects models (LMMs) to account for repeated measurements and variability among participants and stimuli. The results showed that the AVH condition was associated with significantly higher arousal and lower dominance scores than the AV condition. In addition, sadness and fear intensities increased significantly under vibrotactile feedback. Significant interaction effects were observed for valence in sadness, liking in fun and sadness, and anger intensity in fear, indicating that seat-integrated vibrotactile feedback can modulate affective experiences depending on audiovisual content.
To support die-motion design for double-sided symmetric forming of tubular stabilizer bars, this study developed an integrated finite-element (FE)–artificial neural network (ANN) framework for predicting springback and cross-sectional deformation. An R45 Abaqus model was calibrated against tensile responses from three circumferential positions. Free-end simulations established a neutral material inflow of 7.32 mm per 10° of bending. In the fixed-end database analyses, the bending die and support block were translated together by the prescribed axial displacement. Twelve commanded angles and five die-feed offsets generated a 60-case database. After min-max normalization, the 48 development cases underwent five-fold cross-validation, and the 12 holdout cases were reserved for one-time final evaluation. Across the four responses, mean cross-validation R² values ranged from 0.849 to 0.967, and holdout R² values from 0.913 to 0.976. The 2-20-20-4 ANN was then retrained using all 60 cases. R45 fixed-end FE results under the neutral die-feed condition yielded a springback RMSE of 0.275° and a MAPE of 6.29%. Over 10–120°, the maximum relative deviations were 3.52% (minimum outer diameter), 3.65% (minimum wall thickness), and 3.10% (maximum wall thickness). The final surrogate model supports rapid response estimation and angle compensation for new input combinations within the FE design space.
This paper describes an AI-based method for detecting and measuring cracks in the lower mold of a press. During continuous press operation, cracks can form in the lower mold and lead to product defects, which makes manual inspection by operators necessary. In this study, images of the lower mold were captured using a mechanical system developed in our previous study, in which an endoscopic camera is installed at the center of the upper mold. An AI program was then developed using YOLOv8. Images of cracked lower molds were captured at various camera heights, and the crack regions in the captured images were annotated with bounding boxes to train the AI model. Using the trained model, the AI program determines the presence of cracks in the lower mold and calculates the actual size represented by each pixel according to the camera height, thereby enabling the crack dimensions to be measured. The developed program can reduce the burden of manual inspection and provide quantitative information for determining maintenance schedules, thus contributing to the prevention of product defects and the stable operation of press processes.
This study proposes a collaborative robot-based gait simulation system that reproduces prosthetic knee motion and enables quantitative evaluation of control algorithm performance under controlled conditions. The system generates repetitive gait motions from input hip flexion–extension trajectories derived from normative gait data. To validate the system, we evaluated its repeatability, kinematic reproducibility, and control response. Repeatability was assessed by applying identical input trajectories multiple times and quantifying the resulting output variability in terms of the root mean square error (RMSE) and standard deviation. Kinematic reproducibility was evaluated by comparing the simulated knee joint motion of a passive prosthesis with reference gait data from amputees, based on RMSE values and correlation coefficients. In addition, the response characteristics of a microprocessor-controlled prosthetic knee (MPK) were analyzed under repeated input conditions, and the MPK exhibited consistent Lock/Unlock state transitions throughout the gait cycle. The results demonstrate that the proposed system achieves high repeatability and reliably reproduces both the kinematic behavior and the control responses of prosthetic knees. Therefore, the system can serve as an objective and controlled platform for the preclinical evaluation of prosthetic knee control algorithms.
First
Prev
