Frontiers in Robotics and AI最新文献

Autonomous driving has the potential to enhance driving comfort and accessibility, reduce accidents, and improve road safety, with vision sensors playing a key role in enabling vehicle autonomy. Among existing sensors, event-based cameras offer advantages such as a high dynamic range, low power consumption, and enhanced motion detection capabilities compared to traditional frame-based cameras. However, their sparse and asynchronous data present unique processing challenges that require specialized algorithms and hardware. While some models originally developed for frame-based inputs have been adapted to handle event data, they often fail to fully exploit the distinct properties of this novel data format, primarily due to its fundamental structural differences. As a result, new algorithms, including neuromorphic, have been developed specifically for event data. Many of these models are still in the early stages and often lack the maturity and accuracy of traditional approaches. This survey paper focuses on end-to-end event-based object detection for autonomous driving, covering key aspects such as sensing and processing hardware designs, datasets, and algorithms, including dense, spiking, and graph-based neural networks, along with relevant encoding and pre-processing techniques. In addition, this work highlights the shortcomings in the evaluation practices to ensure fair and meaningful comparisons across different event data processing approaches and hardware platforms. Within the scope of this survey, system-level throughput was evaluated from raw event data to model output on an RTX 4090 24GB GPU for several state-of-the-art models using the GEN1 and 1MP datasets. The study also includes a discussion and outlines potential directions for future research.

Cleaning PV (photovoltaic) panels is essential for a PV station, as dirt or dust reduces the effective irradiation of solar energy and weakens the efficiency of converting solar energy into free electrons. The inconsistent (cleaning efficacy) and unsafe (summarized voltage and current) manual method is a challenge for a PV station. Therefore, this paper develops a cleaning robot with PV detection, path planning, and action control. Firstly, a lightweight Mobile-VIT (Mobile Vision Transformer) model with a Self-Attention mechanism was used to improve YOLOv8 (You Only Look Once v8), resulting in an accuracy of 91.08% and a processing speed of 215 fps (frames per second). Secondly, an A* and a DWA (Dynamic Window Approach) path planning algorithm were improved. The simulation result shows that the time consumption decreased from 1.19 to 0.66 s and the Turn Number decreased from 23 to 10 p (places). Finally, the robot was evaluated and calibrated in both indoor and outdoor environments. The results showed that the algorithm can successfully clean PV arrays without manual control, with the rate increasing by 23% after its implementation. This study supports the maintenance of PV stations and serves as a reference for technical applications of deep learning, computer vision, and robot navigation.

Energy consumption is a key concern in modern industrial facilities. Power peak is also a relevant feature in industrial energy analysis and managment, since the electrical infrastructure must be implemented to provide not only the total consumed energy, but the power peaks. Collaborative robots are gaining popularity due to its flexible use and convenient set up. In this context, a power and energy consumption study of the popular UR10 collaborative robot of Universal Robots is reported in this work. For this, an experiment was conducted to obtain current consumption data from the UR10 API, when performing movements with different loads and parameters. Next, the dependency of the trajectory programming parameters on the power peak, total consumed energy, and time spent per trajectory was analyzed. The results show that the higher the speed limit and acceleration limit, the lower the total energy consumed per trajectory, but the higher the power peak. This behavior represents a trade-off: reducing the consumed energy involves increasing the peak power. Based on the captured data, artificial neural network models were trained to predict the power peak and the total energy consumed by the robot when performing a movement under certain parameters. These models were later used by a genetic optimization algorithm to obtain the best parameters for a given target position, providing the most efficient performance while fulfilling a power peak bound.

This article proposes a method that uses information about modules and desired assembly locations within a large truss structure to create a semantic and fiducial aided graph simultaneous localization and mapping (SF-GraphSLAM) algorithm that is better tailored for use during robotic in-space assembly and servicing operations. This is achieved by first reducing the number of modules using a mixed assembly method vs. a strut-by-strut method. Then, each module is correlated to a visual tag (in this article, an AprilTag) to reduce the number of elements being observed further from the number of sub-struts in that module to a single AprilTag marker. Two tags are required to ensure proper deployment of most deployable modules. Subsequently, we are able to use semantic information about the desired transformation matrix between any two adjacent module AprilTags within the desired assembly structure. For our experimentation, we expanded a factor graph smoothing and mapping model and added the semantic information, looking at the smaller number of landmark AprilTags, with a camera representing the robot for simplicity. The mathematical approach to arrive at this new method is included in this article, as are simulations to test it against the state of the art (SOA) using no structural knowledge. Overall, this research contributes to the SOA for both general SLAM work and, more specifically, to the underdeveloped field of SLAM for in-space assembly and servicing of large truss structures. It is critical to ensure that as a robot is assembling the modules, each module is within the desired tolerances to ensure the final structure is within the design requirements. Being able to build a virtual twin of the truss structure as it is being assembled is a key tent pole in achieving large space structures.

Introduction: Children undergoing medical procedures in paediatric Emergency Departments (EDs) often experience significant pain and distress. Socially Assistive Robots (SARs) offer a promising avenue for delivering distraction and emotional support in these high-pressure environments. This study presents the design, development, and formative evaluation of an AI-enhanced SAR to support children during intravenous insertion (IVI) procedures.

Methods: The robot system was developed through a participatory design process involving healthcare professionals, patients, caregivers, and interdisciplinary research teams. The SAR was designed to autonomously adapt its behaviour to the child's affective state using AI planning and social signal processing. A two-cycle usability study was conducted across two Canadian paediatric EDs, involving 25 children and their caregivers. Feedback was collected through observations, interviews, and system logs.

Results: The SAR was successfully integrated into clinical workflows, with positive responses from children, caregivers, and healthcare providers. Usability testing identified key technical and interaction challenges, which were addressed through iterative refinement. The final system demonstrated robust performance and was deemed ready for a formal randomised controlled trial.

Discussion: This work highlights the importance of co-design, operator control, and environmental adaptability in deploying SARs in clinical settings. Lessons learned from the development and deployment process informed six concrete design guidelines for future SAR implementations in healthcare.

Mini cranes play a pivotal role in construction due to their versatility across numerous scenarios. Recent advancements in Reinforcement Learning (RL) have enabled agents to operate cranes in virtual environments for predetermined tasks, paving the way for future real-world deployment. Traditionally, most RL agents use a squashed Gaussian distribution to select actions. In this study, we investigate a mini-crane scenario that could potentially be fully automated by AI and explore replacing the Gaussian distribution with the Kumaraswamy distribution, a close relative of the Beta distribution, for action stochastic selection. Our results indicate that the Kumaraswamy distribution offers computational advantages while maintaining robust performance, making it an attractive alternative for RL applications in continuous control applications.

In response to the growing need for flexibility in handling complex tasks, research on human-robot collaboration (HRC) has garnered considerable attention. Recent studies on HRC have achieved smooth handover tasks between humans and robots by adaptively responding to human states. Collaboration was further improved by conveying the state of the robot to humans via robotic interactive motion cues. However, in scenarios such as collaborative assembly tasks that require precise positioning, methods relying on motion or forces caused by interactions through the shared object compromise both task accuracy and smoothness, and are therefore not directly applicable. To address this, the present study proposes a method to convey the stiffness of the robot to a human arm during collaborative human-robot assembly tasks in a manner that does not affect the shared object or task, aiming to enhance efficiency and reduce human workload. Sixteen participants performed a collaborative assembly task with a robot, which involved unscrewing, repositioning, and reattaching a part while the robot held and adjusted the position of the part. The experiment examined the effectiveness of the proposed method, in which the robot's stiffness was communicated to a participant's forearm. The independent variable, tested within-subjects, was the stiffness presentation method, with three levels: without the proposed method (no presentation) and with the proposed method (real-time and predictive presentations). The results demonstrated that the proposed method enhanced task efficiency by shortening task completion time, which was associated with lower subjective workload scores.

Despite the growing interest in Artificial Intelligence (AI) for pediatric rehabilitation, family engagement in the technologies design remains limited. Understanding how AI-driven tools align with family needs, caregiving routines, and ethical concerns is crucial for their successful adoption. In this study, we actively involved nine families of children with Cerebral Palsy (CP) in an online participatory design workshop, underscoring both the feasibility and the need of integrating family's perspectives into AI development. Families enthusiastically participated, not only sharing insights but also appreciating the opportunity to contribute to shaping future technologies. Their active engagement challenges the assumption that co-design with families is complex or impractical, highlighting how structured yet flexible methodologies can make such crucial initiatives highly effective. The online format further facilitated participation, allowing families to join the discussion and ensuring a diverse range of perspectives. The workshop's key findings reveal three core priorities for families: 1. AI should adapt to daily caregiving routines rather than impose rigid structures; 2. digital tools should enhance communication and collaboration between families and clinicians, rather than replace human interaction; and 3. AI-driven systems could empower children's autonomy while maintaining parental oversight. Additionally, families raised critical concerns about data privacy, transparency, and the need to preserve empathy in AI-mediated care. Our findings reinforce the urgent need to shift toward family-centered AI design, moving beyond purely technological solutions toward ethically responsible, inclusive innovations. This research not only demonstrates the possibility and success of engaging families in co-design processes but also provides a model for future AI development that genuinely reflects the lived experiences of children and caregivers.

This paper addresses the challenge of enabling robots to autonomously prepare meals by bridging natural language recipe instructions and robotic action execution. We propose a novel methodology leveraging Actionable Knowledge Graphs to map recipe instructions into six core categories of robotic manipulation tasks, termed Action Cores cutting, pouring, mixing, preparing, pick and place, and cook and cool. Each AC is subdivided into Action Groups which represent a specific motion parameterization required for task execution. Using the Recipe1M + dataset (Marín et al., IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43, 187-203), encompassing over one million recipes, we systematically analysed action verbs and matched them to ACs by using direct matching and cosine similarity, achieving a coverage of 76.5%. For the unmatched verbs, we employ a neuro-symbolic approach, matching verbs to existing AGs or generating new action cores utilizing a Large Language Model Our findings highlight the versatility of AKGs in adapting general plans to specific robotic tasks, validated through an experimental application in a meal preparation scenario. This work sets a foundation for adaptive robotic systems capable of performing a wide array of complex culinary tasks with minimal human intervention.