Frontiers in Robotics and AI最新文献

Robotic fleet management systems are increasingly vital for sustainable operations in agriculture, forestry, and other field domains where labor shortages, efficiency, and environmental concerns intersect. We present FORMIGA, a fleet management framework that integrates human operators and autonomous robots into a collaborative ecosystem. FORMIGA combines standardised communication through the Robot Operating System with a user-centered interface for monitoring and intervention, while also leveraging large language models to generate executable task code from natural language prompts. The framework was deployed and validated within the FEROX project, a European initiative addressing sustainable berry harvesting in remote environments. In simulation-based trials, FORMIGA demonstrated adaptive task allocation, reduced operator workload, and faster task completion compared to semi-autonomous control, enabling dynamic labor division between humans and robots. By enhancing productivity, supporting worker safety, and promoting resource-efficient operations, FORMIGA contributes to the economic, and environmental dimensions of sustainability, offering a transferable tool for advancing human-robot collaboration in field robotics.

This paper presents a comparative study of data-driven acceleration techniques for mixed-integer bilinear programs (MIBLPs) applied to robot motion planning. MIBLPs combine discrete decision variables and nonlinear constraints, making them computationally challenging for real-time robotics applications. We investigate two reformulation strategies: (1) converting binary variables into continuous variables with complementarity constraints (MPCC), and (2) converting bilinear constraints into mixed-integer linear constraints using McCormick envelopes (MICP). Using offline computed solutions as datasets, we apply K-nearest neighbor methods to warm-start both reformulations. We experimented with the proposed data-driven MIBLP formulation for motion planning on a linear inverted pendulum with contacts, and planning motion using a single rigid body model with mode transitions and contacts. Our results demonstrate that when sufficient data is available, MICP achieves consistently fast solving speeds that are suitable for real-time computation, while MPCC achieves higher success rates with limited amount of data. Our approach is capable of planning motions for the SCALER robot platform to transition between bipedal and quadrupedal configurations to navigate around obstacles without pre-specified gaits. Code for reproducing our results is available at https://github.com/XuanLin/MIBLP_benchmark.

The path planning capability of autonomous robots in complex environments is crucial for their widespread application in the real world. However, long-term decision-making and sparse reward signals pose significant challenges to traditional reinforcement learning (RL) algorithms. Offline hierarchical reinforcement learning offers an effective approach by decomposing tasks into two stages: high-level subgoal generation and low-level subgoal attainment. Advanced Offline HRL methods, such as Guider and HIQL, typically introduce latent spaces in high-level policies to represent subgoals, thereby handling high-dimensional states and enhancing generalization. However, these approaches require the high-level policy to search and generate sub-objectives within a continuous latent space. This remains a complex and sample-inefficient challenge for policy optimization algorithms-particularly policy gradient-based PPO-often leading to unstable training and slow convergence. To address this core limitation, this paper proposes a novel offline hierarchical PPO framework-LG-H-PPO (Latent Graph-based Hierarchical PPO). The core innovation of LG-H-PPO lies in discretizing the continuous latent space into a structured "latent graph." By transforming high-level planning from challenging "continuous creation" to simple "discrete selection," LG-H-PPO substantially reduces the learning difficulty for the high-level policy. Preliminary experiments on standard D4RL offline navigation benchmarks demonstrate that LG-H-PPO achieves significant advantages over advanced baselines like Guider and HIQL in both convergence speed and final task success rates. The main contribution of this paper is introducing graph structures into latent variable HRL planning. This effectively simplifies the action space for high-level policies, enhancing the training efficiency and stability of offline HRL algorithms for long-sequence navigation tasks. It lays the foundation for future offline HRL research combining latent variable representations with explicit graph planning.

Study design: Prospective study.

Objectives: This study aimed to evaluate the accuracy and safety of robot-assisted anterior transpedicular screw (ATPS) fixation in human cervical spine specimens.

Methods: A spine robotic system was used to implant thirty-six 1.2 mm Kirschner wires (K-wires) into the cervical pedicles (C4-C7) of five human specimens. Accuracy was assessed by comparing the planned trajectories with the actual K-wire positions. The Gertzbein-Robbins classification system (GRS), adapted for cervical pedicles, was used to evaluate accuracy; Grades A and B (<2 mm pedicle breach) were considered clinically acceptable. Secondary metrics included entry point and angle offsets.

Results: Of the 36 K-wires implanted, nine were placed in C4 and C6, 10 in C5, and eight in C7. According to the adapted GRS, 25 placements (69.4%) were Grade A, 10 (27.8%) were Grade B, and one was Grade C, resulting in a 97.2% clinically acceptable placement rate. The mean target offset was 2.29 ± 1.72 mm, the entry offset was 2.47 ± 1.57 mm, and the angle offset was 5.67° ± 3.72°. No significant differences were observed between the left and right sides (p > 0.05).

Conclusion: Robot-assisted ATPS fixation in cervical specimens achieved high accuracy with 97.2% of placements rated clinically acceptable, indicating its technical feasibility and potential utility in anterior cervical procedures.

For individuals with unilateral transtibial amputation, powered ankle-foot prostheses have the potential to reduce the metabolic rate of walking, which could contribute to improvements in mobility and quality of life; however, physiological improvements have not been consistently demonstrated in experimental studies. To improve our understanding of the biomechanical mechanisms that drive metabolic rate outcomes, we used a machine learning approach to model the relationship between multimodal biomechanical factors and the metabolic rate of walking with a powered ankle-foot prosthesis. Our model included 50 features describing spatiotemporal parameters, step-to-step transition work, joint kinematics, muscle activity, ground reaction forces, prosthesis settings, and subject characteristics, and resulted in a pseudo-R² of 0.986. Accumulated local effects plots were used to visualize the direction and magnitude of the relationship between each feature and the metabolic rate of walking. The features with the largest effect on metabolic rate were peak unaffected side ankle inversion angle, leading affected leg positive work during the step-to-step transition, and peak affected knee extension angle. This work furthers our knowledge about the biomechanical and physiological response to powered ankle-foot prosthesis use and could assist in developing new strategies to drive reductions in metabolic rate.

Background: Head-mounted virtual reality (VR) simulations are increasingly explored in healthcare, particularly in patient education, stroke rehabilitation, and surgical training. While VR-based simulation plays a growing role in robotic-assisted surgery (RAS) training, the implications of head-mounted VR in this context remain underexamined.

Method: This prospective, randomised, controlled trial with a single-arm crossover compared two training modalities: a head-mounted VR simulation and a conventional console-based simulation. Participants in the experimental group used head-mounted VR as their primary training method, while the control group trained on a conventional console. Both groups completed a running suture task at baseline, midterm, and final assessments on the surgical console. The primary outcome was the composite score from the final assessment.

Results: Fourteen participants were equally distributed between the two arms. Baseline scores showed no significant differences. A two-way repeated measures ANOVA demonstrated significant overall improvement across assessments (F (1.688, 20.26) = 48.34, p < 0.001, partial η² = 0.80). No statistical difference was found in final composite scores (mean difference: 8.4 ± 9.45, p = 0.391, Cohen's d = -0.48), midterm scores, or granular kinematic data. However, non-inferiority could not be established as the confidence interval fell outside our pre-set margin. The crossover group required less time (mean difference: 39 ± 9.01 min, p = 0.004) and fewer attempts (mean difference: 8 ± 2.2, p = 0.009) to reach benchmark performance compared to the control group.

Conclusion: Both head-mounted VR and console-based training significantly improved fundamental RAS skills in novices. While our study showed that the VR training shortened the time and attempts required to reach proficiency benchmarks, the small scale of this trial and the breadth of the confidence intervals mean the results should be viewed as preliminary observations. These results provide an initial signal of feasibility that warrants confirmation in larger studies.

A cybernetic avatar (CA) is a concept that encompasses not only avatars representing virtual bodies in cyberspace but also information and communication technology (ICT) and robotic technologies that enhance the physical, cognitive, and perceptual capabilities of humans. CAs can enable multiple people to remotely operate numerous avatars and robots together to perform complex tasks on a large scale and create the necessary infrastructure for their operation and other related activities. However, due to the novelty of this concept, the ethical, legal, and social implications (ELSI) of CAs have not been discussed sufficiently. Therefore, the objective of this paper is to provide an overview of ELSI in the context of a CA, taking into account the implications from fields similar to that of CAs, such as robotic avatars, virtual avatars, metaverses, virtual reality, extended reality, social robots, human-robot interaction, telepresence, telexistence, embodied technology, and exoskeletons. In our review of ELSI in related fields, we found common themes: safety and security, data privacy, identity theft and identity loss, manipulation, intellectual property management, user addiction and overdependence, cyber abuse, risk management in a specific domain (e.g., medical applications), regulatory gaps, balance between free expression and harmful content, accountability, transparency, distributive justice, prevention of inequalities, dual use, and conceptual changes of familiarity. These issues should not be ignored when considering the social implementation of CAs.

Wetware Network-Based Artificial Intelligence (WNAI) introduces a new approach to robotic cognition and artificial intelligence: autonomous cognitive agents built from synthetic chemical networks. Rooted in Wetware Neuromorphic Engineering, WNAI shifts the focus of this emerging field from disembodied computation and biological mimicry to reticular chemical self-organization as a substrate for cognition. At the theoretical level, WNAI integrates insights from network cybernetics, autopoietic theory and enaction to frame cognition as a materially grounded, emergent phenomenon. At the heuristic level, WNAI defines its role as complementary to existing leading approaches. On the one hand, it complements embodied AI and xenobotics by expanding the design space of artificial embodied cognition into fully synthetic domains. On the other hand, it engages in mutual exchange with neural network architectures, advancing cross-substrate principles of network-based cognition. At the technological level, WNAI offers a roadmap for implementing chemical neural networks and protocellular agents, with potential applications in robotic systems requiring minimal, adaptive, and substrate-sensitive intelligence. By situating wetware neuromorphic engineering within the broader landscape of robotics and AI, this article outlines a programmatic framework that highlights its potential to expand artificial cognition beyond silicon and biohybrid systems.

As robots became increasingly integrated into daily life, their ability to influence human emotions through verbal and nonverbal expressions is gaining attention. While robots have been explored for their role in emotional expression, their potential in emotion regulation particularly in mitigating or amplifying embarrassment remains under-explored in human-robot interaction. To address this gap, this study investigates whether and how robots can regulate the embarrassment emotion through their responses. A between-subjects experiment was conducted with 96 participants (48 males and 48 females) using the social robot Furhat. Participants experienced an embarrassing situation induced by a failure of meshing scenario, followed by the robot adopting one of three response attitudes: neutral, empathic, or ridiculing. Additionally, the robot's social agency was manipulated by varying its facial appearance between a human-like and an anime-like appearances. The findings indicate that embarrassment was effectively induced, as evidenced by physiological data, body movements, facial expressions, and participants' verbal responses. The anime-faced robot elicited lower embarrassment and arousal due to its lower perceived social agency and anthropomorphism. The robot's attitude was the dominant factor shaping participants' emotional responses and perceptions. The neutral and empathic attitudes with an anime face were found to be the most effective in eliciting mild emotions and mitigating embarrassment. Interestingly, an empathic attitude is suspected to be favored over a neutral one as it elicited the lowest embarrassment. However, an empathic attitude risks shaming the participant due to its indirect confrontation that inherently acknowledges the embarrassing incident which is undesirable in Japanese culture. Nevertheless, in terms of the robot's perceived evaluation by participants, a neutral attitude was the most favored. This study highlights the role of robot responses in emotion regulation, particularly in embarrassment control, and provides insights into designing socially intelligent robots that can modulate human emotions effectively.

Collaboration between humans and robots is essential for optimizing the performance of complex tasks in industrial environments, reducing worker strain, and improving safety. This paper presents an integrated human-robot collaboration (HRC) system that leverages advanced intention recognition for real-time task sharing and interaction. By utilizing state-of-the-art human pose estimation combined with deep learning models, we developed a robust framework for detecting and predicting worker intentions. Specifically, we employed LSTM-based and transformer-based neural networks with convolutional and pooling layers to classify human hand trajectories, achieving higher accuracy compared to previous approaches. Additionally, our system integrates dynamic movement primitives (DMPs) for smooth robot motion transitions, collision prevention, and automatic motion onset/cessation detection. We validated the system in a real-world industrial assembly task, demonstrating its effectiveness in enhancing the fluency, safety, and efficiency of human-robot collaboration. The proposed method shows promise in improving real-time decision-making in collaborative environments, offering a safer and more intuitive interaction between humans and robots.