Editorial for the Special Issue “Robotic Intelligence Development of AI in Robot Perception, Learning, and Decision”

Artificial intelligence is increasingly reshaping the foundations of robotics by enabling machines to perceive complex environments, learn from interaction, make adaptive decisions, and collaborate with humans in a safe and effective manner [1]. Recent advances in deep learning, reinforcement learning, semantic perception, intelligent control, and human–robot interaction have accelerated the transition of robotic systems from pre-programmed machines toward autonomous, context-aware, and adaptive agents. These developments are particularly important for applications in mobile manipulation, industrial automation, medical robotics, smart manufacturing, educational robotics, and collaborative workspaces.

This Special Issue, “Robotic Intelligence Development of AI in Robot Perception, Learning, and Decision”, was launched to provide a platform for recent research on the integration of AI techniques into robotic perception, learning, control, and decision-making systems. The contributions collected in this Special Issue reflect the diversity and vitality of this field. They address both fundamental methodological challenges and practical application scenarios, including robust visual SLAM, semantic mapping, reinforcement learning-based control, human–robot collaboration, robotic calibration, medical robotic imaging, mobile manipulation, and educational robotics.

The first group of papers focuses on robot perception, mapping, and scene understanding. Liu et al. (list No. 1) proposed LDFE-SLAM, a light-aware deep front-end for robust visual SLAM under challenging illumination conditions. By integrating illumination-adaptive enhancement, SuperPoint-based deep feature detection, and LightGlue-based matching, the proposed system demonstrated improved robustness in low-light environments, where conventional visual SLAM methods often suffer from degraded feature extraction and unstable tracking. This work highlights the importance of co-designing visual enhancement, feature extraction, and learned matching modules for reliable robotic localization in realistic environments.

Alotaibi et al. present a deep learning-driven semantic mapping strategy for robotic inspection of desalination facilities. Their system integrates LiDAR, RGB-D, and odometry data to construct semantic maps capable of detecting and labeling critical infrastructure components. Through simulations and real-world experiments using an ROS-based robotic platform, the study demonstrates the potential of semantic mapping for autonomous inspection in industrial facilities. Together, these perception-orientated contributions show how AI can support robots in moving beyond geometric mapping toward richer environmental understanding.

A second group of papers addresses learning-based control and decision-making. Ma et al. introduce a dual-arm fabric-flattening method based on hybrid imitation and reinforcement learning. Their cascaded Proposal–Action network was first trained using human demonstrations and then refined through reinforcement learning with real-world feedback. The method achieved high success rates and generalized effectively to fabrics with different physical properties. This work is a representative example of how imitation learning and reinforcement learning can be combined to address robotic manipulation tasks involving deformable objects, which remain challenging due to nonlinear dynamics and uncertainty.

Mon proposes a decoupled reinforcement hybrid PPO–sliding control strategy for underactuated systems, with applications to Cart–Pole and Acrobot benchmarks. By integrating policy optimization with sliding mode control, the study aimed to improve stability, robustness, and learning efficiency in nonlinear control problems. Although some limitations remain in sustained Acrobot balance control, the work contributes to the broader effort of combining model-free reinforcement learning with classical control principles for safer and more stable robotic decision-making.

Another major theme of this Special Issue is human–robot collaboration and human-centered robotic intelligence. Moghaddam and Arrichiello propose an adaptive variable admittance control framework for intent-aware human–robot collaboration. Their approach introduces an intent-aware human force generator and extends variable admittance control by incorporating adaptive stiffness alongside damping and inertia. Using online adaptation through a self-supervised learning mechanism, the framework was evaluated in a dual-arm collaborative manipulation scenario. This study provides an important tool for stress-testing collaborative controllers under dynamic and unpredictable human intentions.

Zhou et al. developed a methodology for human–robot collaborative assembly based on human skill imitation and learning. Their framework uses electromyography signals as an interaction interface and combines human skill imitation with reinforcement learning. A temporal graph neural network with angle-guided attention was designed for dynamic arm force estimation, while an expert reward function and fuzzy experience replay were introduced to improve collaborative comfort, smoothness, and assembly accuracy. This contribution demonstrates the potential of combining bio-signal-based intention estimation, graph learning, and reinforcement learning for more natural and effective human–robot collaboration.

Several papers in this Special Issue further address precision, robustness, and intelligent compensation in robotic systems. Zhu et al. propose an integrated framework combining a semi-active magnetorheological elastomer intelligent isolation system with an active trajectory tracking controller for mobile manipulators operating under continuous impact disturbances. By coupling vibration isolation with disturbance-observer-based control, the study demonstrates improved trajectory tracking performance under harsh excitation conditions. This work illustrates how intelligent mechanical design and advanced control can be jointly optimized to enhance robot robustness.

Liu et al. (list No. 8) developed a two-stage compensation framework for improving robotic positioning accuracy. Their method first uses a memory-based red-billed blue magpie optimizer to identify and compensate for geometric errors and then applies an adaptive momentum particle swarm optimization-tuned graph neural network to model non-geometric errors. The framework was validated on a six-axis industrial robot and showed improved residual error compensation. This study reflects a growing trend toward combining evolutionary optimization, graph neural networks, and calibration strategies for precision-critical robotic applications.

Medical and application-specific robotic intelligence is also represented in this Special Issue. Ward et al. present an autonomous fluoroscopic imaging system for catheter insertions based on a bilateral control scheme. Their simulation study coupled fluoroscopic image feedback with the kinematics of a tendon-driven robotic catheter and a motorized C-arm. By maintaining optimal imaging geometry during autonomous catheter insertion, the system aims to support safer and more effective image-guided procedures. Although further validation under realistic imaging noise and hardware-in-the-loop conditions is needed, their work provides a promising foundation for intelligent medical robotic systems.

Finally, Plókai et al. report on the deployment of an educational mobile robot, focusing on software tools for processing, analyzing, and visualizing sensor data. Using odometry and inertial measurement data from the PlatypOUs mobile robot platform, the study applied moving averages, correlation analysis, and exponential smoothing to improve data interpretability and reveal motion dynamics. This contribution emphasizes the educational and methodological value of data-driven analysis in mobile robotics and STEM learning environments.

Taken together, the papers in this Special Issue demonstrate that robotic intelligence is not defined by a single algorithmic paradigm. Instead, it emerges from the integration of perception, learning, control, optimization, interaction, and domain knowledge. Several cross-cutting trends can be identified. Firstly, deep learning is becoming central to robot perception, particularly in visual SLAM, semantic mapping, and sensor fusion. Secondly, reinforcement learning and imitation learning are increasingly being combined with classical control methods to improve stability, data efficiency, and task performance. Thirdly, human-centered robotics is moving toward richer representations of human intention, skill, comfort, and safety. Fourthly, intelligent compensation and adaptive control are playing a growing role in improving the precision and robustness of robots operating in uncertain, dynamic, or physically challenging environments.

Despite these advances, important challenges remain. Many AI-enabled robotic systems still face difficulties in generalizing from simulation to real-world deployment, maintaining safety under distribution shifts, and operating reliably with limited or noisy data. Future research should further explore embodied intelligence to physical AI [2], Large Language Model-assisted task planning [3], robotic surgery [4], brain-inspired neuromorphic devices [5], and brain–computer interfaces [6]. In addition, closer integration between mechanical design, sensing hardware, control theory, and learning algorithms will be essential in building next-generation robotic systems that are not only intelligent but also safe, efficient, and deployable.

In conclusion, this Special Issue features a timely collection of research contributions that advance the development of AI in robot perception, learning, control, and decision-making. The published papers collectively demonstrate how intelligent robotic systems can perceive complex environments, learn from data and interaction, adapt to uncertainty, collaborate with humans, and perform demanding tasks across industrial, medical, educational, and service domains. We hope that this Special Issue will stimulate further research and collaboration in the rapidly evolving field of robotic intelligence.

The Guest Editors would like to express their sincere gratitude to all authors who contributed their valuable work to this Special Issue. We also thank the reviewers for their constructive comments and careful evaluations, which helped improve the quality of the published papers. Finally, we acknowledge the editorial team of Machines for their professional support throughout the preparation and publication of this Special Issue.