Artificial Intelligence and Its Applications in Robotics

1. Introduction

Prior to the advent of intelligent robots, artificial intelligence (AI) and robotics were considered two distinct fields; however, the rapid development of hardware with higher computational power and more precise sensors, along with the exponential increase in data, has contributed to significant advancements in the integration of the two fields [1]. Moreover, as part of efforts to enhance quality of life, there has been a growing demand from industries to make robots more versatile and cognitively capable in complex environments—achievements that can be accomplished through the deployment of AI methods [2]. Some of the aspects of robotic systems that AI has significantly improved include object detection and recognition, speech and gesture recognition, motion planning, control, and localization [3]. These advancements have made robots more adaptable, accurate, reliable, and autonomous.

Beyond autonomous driving and vehicles, intelligent robotic systems are increasingly present across various domains. In healthcare, for instance, surgical robots, remote patient monitoring, precision medicine, and early diagnostics are major outcomes of the fusion of AI and robotics [4]. Service robots are already being used in restaurants, hospitals, and the tourism industry to assist customers [5]; AI-powered robots are being used in universities to enhance the quality of education by improving the online learning experience and increasing both the quality and quantity of resources available to both students and faculty [6]; and intelligent robots are widely applied in the renewable energy sector to reduce operation and maintenance costs while increasing worker safety [7].

This Special Issue delivers ten outstanding research articles that showcase cutting-edge AI methods for various robotic applications, and one review article on the implementation of robotics operating system (ROS)-based reinforcement learning (RL). Collectively, these articles demonstrate how AI technologies can broaden the spectrum of robotic applications across multiple fields. The contributions in this issue can be categorized into three main areas: (1) robot design, (2) AI-based control, and (3) AI-based perception.

2. An Overview of the Published Articles

Designing robots to perform specific tasks using AI techniques is essential. In order to achieve this goal, appropriate sensors, computing devices, and actuators must be selected and the kinematics of the robot must be carefully designed. A small mobile robot was designed by Wong et al. (Contribution 3), consisting of a computer, providing obstacle avoidance capability, as well as a 2D LiDAR and two fisheye cameras to recognize signs and detect pathways. To achieve the desired capabilities, YOLOv4-tiny Canny edge detection and a distance-based filtering mechanism were implemented. The Miyazaki Vermin Repulsion Robot (MiVeReR) was designed by Lee et al. (Contribution 9) to protect livestock against infections. It generates adjustable acousto-optic signals to drive away wild animals and provides a monitoring capability to track their locations. Both indoor and outdoor experiments were conducted to verify its effectiveness and reliability.

As AI technologies improve, robots have now come to be capable of performing sophisticated tasks that were not previously possible with classical control and planning schemes. Herman (Contribution 1) designed a new control scheme that combines a backstepping method with adaptive integral sliding mode control to perform trajectory tracking for an unmanned underwater vehicle. To enhance performance, adaptive laws were implemented to mitigate the effects of parameter perturbations and external disturbances. In turn, Parsons et al. (Contribution 4) proposed a unique path planning algorithm and verified it by coupling it to an unmanned ground vehicle (UGV). The ant colony optimization (ACO) algorithm was adopted to solve the covering salesman problem (CSP), and the genetic algorithm (GA) was then used to reduce the total path length. Finally, to avoid sharp turning angles, the corner-cutting and Reeds–Shepp methods were implemented. A minimal neural network (NN)-based model-free control structure was designed by Toussaint and Raison (Contribution 6) to achieve precise, high-speed trajectory tracking for robot manipulators. They trained and compared two NN architectures—a separate NN for each joint (1NN) and a single global NN (GNN)—producing modified position commands that can be applied to a proportional–derivative (PD) controller, reducing tracking errors. Zhang et al. (Contribution 8) proposed a dual-layer reinforcement learning (RL) algorithm to control the locomotion of a quadruped robot. Their deep double Q-network (DDQN) generates an appropriate walking speed based on observations, and their proximal policy optimization (PPO) algorithm then executes the action at the assigned speed. The proposed model was verified in the Isaac Gym environment. An extensive review of ROS-based RL implementations was conducted by Aljamal et al. (Contribution 11), providing a summary and analysis of such applications across multiple domains, including motion planning, path planning, robot control, collision avoidance, and navigation, across various robotic platforms such as industrial robots, mobile robots, and drones.

Beyond control algorithms, AI can supply rich information that robots can use when performing various tasks. Song et al. (Contribution 2) developed a novel ML model called U-TFF, which compiles time–frequency fusion (TFF) blocks into a U-shaped architecture. The energy consumption measurements of the robot’s manipulator were transformed into the frequency domain to train U-TFF, and the reconstruction error of this semi-supervised model was used to detect anomalies, achieving an accuracy of 0.93. Sun et al. (Contribution 5) presented a refined prior-guided, category-level 6D pose estimation model. They introduced a refinement module and an attention mechanism to enhance the robustness of the network and address the intra-class variation problem. The proposed method estimated the 6D pose of objects under partial occlusion and various lighting conditions. Zbiss et al. (Contribution 7) developed a framework to find the optimal base placement for a multi-robot vehicle painting system. They constructed two cost functions to maximize paint coverage and minimize collision, which were then incorporated into a multi-objective optimization problem without the need for trajectory planners. Zuo and Zhang (Contribution 10) developed an audio-guided space-time memory network (AG-STMNet) to build a lightweight audiovisual segmentation (AVS) framework. The mask generator identifies the sounding objects, and the AG-STMNet tracks them while mitigating the effects of environmental noise. The framework achieved an mIoU score of 41.5 on the AVSBench dataset.

3. Conclusions

This Special Issue spotlights various AI technologies developed for robotic systems, aiming to provide solutions to existing limitations and challenges across the field. These efforts are extremely valuable in terms of expanding the robotics community and guiding us one step closer to AI-integrated lifestyles. The overarching goal of this Special Issue is to introduce AI tools such as RL, ML, adaptive control, pose estimation, path planning, randomized optimization, object detection and AVS, and to explore their integration into robotic platforms to enhance the robots’ perception and control. The algorithms presented in this Special Issue can serve as cornerstones for the development of more robust and intelligent robotic systems that will become integral elements of our future.