Frontiers in Robotics and AI最新文献

Animal muscles have complex, three-dimensional structures with fibers oriented in various directions. The tongue, in particular, features a highly intricate muscular system composed of four intrinsic muscles and several types of extrinsic muscles, enabling flexible and diverse movements essential for feeding, swallowing, and speech production. Replicating these structures could lead to the development of multifunctional manipulators and advanced platforms for studying muscle-motion relationships. In this study, we developed a pig tongue soft robot that focuses on replicating the intrinsic muscles using thin McKibben artificial muscles, silicone rubber, and gel. We began by performing three-dimensional scans and sectional observations in the coronal and sagittal planes to examine the arrangement and orientation of the intrinsic muscles in the actual pig tongue. Additionally, we used the diffusible iodine-based contrast-enhanced computed tomography (Dice-CT) technique to observe the three-dimensional flow of muscle pathways. Based on these observations, we constructed a three-dimensional model and molded the pig tongue shape with silicone rubber and gel, embedding artificial muscles into the robot body. We conducted experiments to assess both the motion of the tongue robot's tip and its stiffness during muscle contractions. The results confirmed characteristic tongue motions, such as tip extension, flexion, and lateral bending, as well as stiffness changes during actuation, suggesting the potential for this soft robot to serve as a platform for academic and engineering studies.

Introduction: Subsea applications recently received increasing attention due to the global expansion of offshore energy, seabed infrastructure, and maritime activities; complex inspection, maintenance, and repair tasks in this domain are regularly solved with pilot-controlled, tethered remote-operated vehicles to reduce the use of human divers. However, collecting and precisely labeling submerged data is challenging due to uncontrollable and harsh environmental factors. As an alternative, synthetic environments offer cost-effective, controlled alternatives to real-world operations, with access to detailed ground-truth data. This study investigates the potential of synthetic underwater environments to offer cost-effective, controlled alternatives to real-world operations, by rendering detailed labeled datasets and their application to machine-learning.

Methods: Two synthetic datasets with over 1000 rendered images each were used to train DeepLabV3+ neural networks with an Xception backbone. The dataset includes environmental classes like seawater and seafloor, offshore structures components, ship hulls, and several marine growth classes. The machine-learning models were trained using transfer learning and data augmentation techniques.

Results: Testing showed high accuracy in segmenting synthetic images. In contrast, testing on real-world imagery yielded promising results for two out of three of the studied cases, though challenges in distinguishing some classes persist.

Discussion: This study demonstrates the efficiency of synthetic environments for training subsea machine learning models but also highlights some important limitations in certain cases. Improvements can be pursued by introducing layered species into synthetic environments and improving real-world optical information quality-better color representation, reduced compression artifacts, and minimized motion blur-are key focus areas. Future work involves more extensive validation with expert-labeled datasets to validate and enhance real-world application accuracy.

Object pose estimation is essential for computer vision applications such as quality inspection, robotic bin picking, and warehouse logistics. However, this task often requires expensive equipment such as 3D cameras or Lidar sensors, as well as significant computational resources. Many state-of-the-art methods for 6D pose estimation depend on deep neural networks, which are computationally demanding and require GPUs for real-time performance. Moreover, they usually involve the collection and labeling of large training datasets, which is costly and time-consuming. In this study, we propose a template-based matching algorithm that utilizes a novel perceptual hashing method for binary images, enabling fast and robust pose estimation. This approach allows the automatic preselection of a subset of templates, significantly reducing inference time while maintaining similar accuracy. Our solution runs efficiently on multiple devices without GPU support, offering reduced runtime and high accuracy on cost-effective hardware. We benchmarked our proposed approach on a body-in-white automotive part and a widely used publicly available dataset. Our set of experiments on a synthetically generated dataset reveals a trade-off between accuracy and computation time superior to a previous work on the same automotive-production use case. Additionally, our algorithm efficiently utilizes all CPU cores and includes adjustable parameters for balancing computation time and accuracy, making it suitable for a wide range of applications where hardware cost and power efficiency are critical. For instance, with a rotation step of 10° in the template database, we achieve an average rotation error of $10°$ , matching the template quantization level, and an average translation error of 14% of the object's size, with an average processing time of $0.3s$ per image on a small form-factor NVIDIA AGX Orin device. We also evaluate robustness under partial occlusions (up to 10% occlusion) and noisy inputs (signal-to-noise ratios [SNRs] up to 10 dB), with only minor losses in accuracy. Additionally, we compare our method to state-of-the-art deep learning models on a public dataset. Although our algorithm does not outperform them in absolute accuracy, it provides a more favorable trade-off between accuracy and processing time, which is especially relevant to applications using resource-constrained devices.

Introduction: This study focused on the psychological evaluation of an avatar robot in two distinct regions, Dubai in the Middle East and Japan in the Far East. Dubai has experienced remarkable development in advanced technology, while Japan boasts a culture that embraces robotics. These regions are distinctively characterized by their respective relationships with robotics. In addition, the use of robots as avatars is anticipated to increase, and this research aimed to compare the psychological impressions of people from these regions when interacting with an avatar as opposed to a human.

Methods: Considering that avatars can be presented on screens or as physical robots, two methodologies were employed: a video presentation survey (Study 1, Dubai: n = 120, Japan: n = 120) and an experiment involving live interactions with a physical robot avatar (Study 2, Dubai: n = 28, Japan: n = 30).

Results and discussion: Results from the video presentations indicated that participants from Dubai experienced significantly lower levels of discomfort towards the avatar compared to their Japanese counterparts. In contrast, during live interactions, Japanese participants showed a notably positive evaluation towards a Japanese human operator. The findings suggest that screen-presented avatars may be more readily accepted in Dubai, while humans were generally preferred over avatars in terms of positive evaluations when physical robots were used as avatars. The study also discusses the implications of these findings for the appropriate tasks for avatars and the relationship between cultural backgrounds and avatar evaluations.

Laboratory automation requires reliable and precise handling of microplates, but existing robotic systems often struggle to achieve this, particularly when navigating around the dynamic and variable nature of laboratory environments. This work introduces a novel method integrating simultaneous localization and mapping (SLAM), computer vision, and tactile feedback for the precise and autonomous placement of microplates. Implemented on a bi-manual mobile robot, the method achieves fine-positioning accuracies of $\pm$ 1.2 mm and $\pm$ 0.4°. The approach was validated through experiments using both mockup and real laboratory instruments, demonstrating at least a 95% success rate across varied conditions and robust performance in a multi-stage protocol. Compared to existing methods, our framework effectively generalizes to different instruments without compromising efficiency. These findings highlight the potential for enhanced robotic manipulation in laboratory automation, paving the way for more reliable and reproducible experimental workflows.

After the COVID-19 pandemic, the adoption of distance learning has been accelerated in educational institutions in multiple countries. In addition to using a videoconferencing system with camera images, avatars can also be used for remote classes. In particular, an android avatar with a sense of presence has the potential to provide higher quality education than a video-recorded lecture. To investigate the specific educational effects of android avatars, we used a Geminoid. an android with the appearance of a specific individual, and conducted both laboratory experiment and large-scale field experiment. The first compared the android avatar lecture with a videoconferencing system. We found that the use of an android avatar for the lecture led to the significantly higher subjective feelings of being seen, feeling more motivated, and focused on the lecture compared to the video lecture. We further conducted a large-scale field experiment with an android avatar to clarify what contributes to such educational effects. The results suggest that the students' perception of android's anthroppomorphism and competence has a positive impact, and discomfort has a negative impact on the subjective experence of educational effect. These results indicate the role of embodied anthropomorphization in positive educational experience. The important point of this study is that both the laboratory experiment and the large-scale experiment were conducted to clarify the educational effects of androids. These results support several related studies and are clarified in detail. Based on these results, the potential for the future usage of androids in education is discussed.

This paper explores the applicability of bicycle-inspired balance control in a quadruped robot model. Bicycles maintain stability and change direction by intuitively steering the handle, which induces yaw motion in the body frame and generates an inertial effect to support balance. Inspired by this balancing strategy, we implemented a similar mechanism in a quadruped robot model, introducing a yaw trunk joint analogous to a bicycle's steering handle. Simulation results demonstrate that the proposed model achieves stable high-speed locomotion with robustness against external disturbances and maneuverability that allows directional changes with only slight speed reduction. These findings suggest that utilizing centrifugal force plays a critical role in agile locomotion, aligning with the movement strategies of cursorial animals. This study underscores the potential of bicycle balance control as an effective and straightforward control approach for enhancing the agility and stability of quadruped robots as well as potentially offering insights into animal motor control mechanisms for agile locomotion.

Reliable proprioception and feedback from soft sensors are crucial for enabling soft robots to function intelligently in real-world environments. Nevertheless, soft sensors are fragile and are susceptible to various damage sources in such environments. Some researchers have utilized redundant configuration, where healthy sensors compensate instantaneously for lost ones to maintain proprioception accuracy. However, achieving consistently reliable proprioception under diverse sensor degradation remains a challenge. This paper proposes a novel framework for graceful degradation in redundant soft sensor systems, incorporating a stochastic Long Short-Term Memory (LSTM) and a Time-Delay Feedforward Neural Network (TDFNN). The LSTM estimates readings from healthy sensors to compare them with actual data. Then, statistically abnormal readings are zeroed out. The TDFNN receives the processed sensor readings to perform proprioception. Simulation experiments with a musculoskeletal leg that contains 40 nonlinear soft sensors demonstrate the effectiveness of the proposed framework. Results show that the knee angle proprioception accuracy is retained across four distinct degradation scenarios. Notably, the mean proprioception error increases by less than 1.91°(1.36%) when $30\%$ of the sensors are degraded. These results suggest that the proposed framework enhances the reliability of soft sensor proprioception, thereby improving the robustness of soft robots in real-world applications.

We present WearMoCap, an open-source library to track the human pose from smartwatch sensor data and leveraging pose predictions for ubiquitous robot control. WearMoCap operates in three modes: 1) a Watch Only mode, which uses a smartwatch only, 2) a novel Upper Arm mode, which utilizes the smartphone strapped onto the upper arm and 3) a Pocket mode, which determines body orientation from a smartphone in any pocket. We evaluate all modes on large-scale datasets consisting of recordings from up to 8 human subjects using a range of consumer-grade devices. Further, we discuss real-robot applications of underlying works and evaluate WearMoCap in handover and teleoperation tasks, resulting in performances that are within 2 cm of the accuracy of the gold-standard motion capture system. Our Upper Arm mode provides the most accurate wrist position estimates with a Root Mean Squared prediction error of 6.79 cm. To evaluate WearMoCap in more scenarios and investigate strategies to mitigate sensor drift, we publish the WearMoCap system with thorough documentation as open source. The system is designed to foster future research in smartwatch-based motion capture for robotics applications where ubiquity matters. www.github.com/wearable-motion-capture.

When developing general-purpose robot software components, we often lack complete knowledge of the specific contexts in which they will be executed. This limits our ability to make predictions, including our ability to detect program bugs statically. Since running a robot is an expensive task, finding errors at runtime can prolong the debugging loop or even cause safety hazards. This paper proposes an approach to help developers catch these errors as soon as we have some context (typically at pre-launch time) with minimal additional efforts. We use embedded domain-specific language (DSL) techniques to enforce early checks. We describe design patterns suitable for robot programming and show how to use these design patterns for DSL embedding in Python, using two case studies on an open-source robot skill platform SkiROS2, designed for the composition of robot skills. These two case studies help us understand how to use DSL embedding on two abstraction levels: the high-level skill description that focuses on what the robot can do and under what circumstances and the lower-level decision-making and execution flow of tasks. Using our DSL EzSkiROS, we show how our design patterns enable robotics software platforms to detect bugs in the high-level contracts between the robot's capabilities and the robot's understanding of the world. We also apply the same techniques to detect bugs in the lower-level implementation code, such as writing behavior trees (BTs), to control the robot's behavior based on its capabilities. We perform consistency checks during the code deployment phase, significantly earlier than the typical runtime checks. This enhances the overall safety by identifying potential issues with the skill execution before they can impact robot behavior. An initial study with SkiROS2 developers shows that our DSL-based approach is useful for finding bugs early and thus improving the maintainability of the code.