Astrictive-type grippers, which generate gripping forces from adhesive forces at the contact surface, such as suction cup, are popular end-effectors as picking solutions because of their simplicity and small working space. However, the adhesive force of the astrictive gripper decreases with increasing complexity of the object surface; thus, its application has been restricted to simple picking of objects with a flat surface. Here, in this article, we present an all-round honeycomb astrictive gripper that has an orthotropic surface tension for grasping highly irregular shaped objects with an uneven surface. The design is inspired by mimicking the two-level (macro- and mesoscale) shape adaptation of the octopus's leg. The stiffness-variable structure is also consisted to change its stiffness similar to the function of octopus's leg, and owing to the combination of these structures makes possible to perform various tasks, including hammering, breakfast serving, and vaccination, which were not possible for previous astrictive gripper.

Handheld virtual reality (VR) controllers are necessary for creating immersive experiences. In this paper, we propose a gated RNN-based sequence model that estimates the joint torques of a serially linked handheld VR system interface from a sequential position input. In our previous study, we proposed a motion planning algorithm for articulated systems based on the active contour model that optimizes the positions of each joint torque based on the measured base position (6-Degrees of Freedom). Because the position-toposition scheme, which calculates the joint positions from a given base position, illustrated several limitations concerning safety (i.e. unable to handle unexpected contact with the surroundings), our current study proposes a position-to-torque generation scheme that estimates the joint torques from the measured base position sequences. To that end, we trained the sequences of joint torques and the sequence of the 6-DoF base position as a supervised learning task. To model the multivariate temporal information of the sequences, we employed a gated recurrent unit. The experimental results validate the successful generation of joint trajectory profiles.

Despite the rapid expansion of EVs (Electric Vehicles), current EV charging systems remain inconvenient for users. To address this issue, robot-assisted EV charging systems are emerging as promising solutions, requiring precise EV inlet localization as a key technology. However, existing vision-based methods face fundamental trade-offs: 2D RGB provides dense visual information but suffers from illumination sensitivity and depth ambiguity, while 3D PC (Point Cloud) provides geometric precision with limitations in sparse structures and high computational complexity. To overcome these challenges, we propose MDSNet (Multimodal Dual Stream Network), a novel framework that combines 2D and 3D CNNs through two key innovations: (1) modality-specific axial attention mechanisms for efficient feature fusion, and (2) hierarchical data slicing, tailored to each modality-specific characteristic. Unlike single-modality approaches, MDSNet effectively leverages complementary strengths while mitigating individual weaknesses of both RGB and PC data. The proposed method is evaluated on RGB-PC data pairs acquired from realistic EV scenarios under various real-world environments including multiple distances, viewing angles, and illumination variations. Experimental results demonstrate state-of-the-art performance, establishing MDSNet as a practical solution for autonomous robot-assisted EV charging systems.

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.