Elevating Mobile Robotics: Pioneering Applications of Artificial Intelligence and Machine Learning

One of the most widely used approaches to mobile robot control is subsumption architecture, which is often used in conjunction with AI techniques such as machine learning and neural networks. Subsumption architecture is a behavior-based approach that involves the use of reactive controllers organized into a hierarchy of layers, where each layer is responsible for a specific behavior. The layers are arranged in a way that allows higher-level behaviors to subsume lower-level behaviors, hence the name "subsumption architecture."

The use of subsumption architecture in mobile robot control has several advantages. First, it allows for the creation of complex behaviors by combining simple behaviors. Second, it enables the robot to react quickly to changes in the environment, as the reactive controllers are designed to respond to sensor feedback in real-time. Finally, it allows for the creation of robust and fault-tolerant systems, as the layers can be designed to operate independently of each other.

The integration of subsumption architecture with AI techniques such as machine learning and neural networks has led to significant advancements in mobile robot control. For example, machine learning algorithms can be used to train the higher-level behaviors of a mobile robot, while subsumption architecture can be used to implement the low-level behaviors. This combination of techniques allows for the creation of highly adaptive and intelligent mobile robots that can perform complex tasks and make autonomous decisions. In this study, we will explore the current landscape of AI applications in the control of mobile robots, with a particular emphasis on machine learning and neural networks. We will also examine the role of subsumption architecture in mobile robot control and its integration with AI techniques. By the end of this study, you will have a comprehensive understanding of the cutting-edge applications of AI and machine learning in mobile robotics and the role of subsumption architecture in creating highly adaptive and intelligent mobile robots. For moving a mobile robot to a goal position, it is told to execute a particular type of force into the wheels’ motor for a period of time rather than going directly to the goal position coordinates. The robot’s movement is programmed using a controller, a piece of software that tells the robot what to do after sensing the environment until a control objective is achieved. Reactive controls are oriented to deal with simple reflex agents. Most of the actions a robot can execute are separable in subcomponents [26]. These subcomponents can be modelled as finite-state machines that react to sensor feedback [27]. Subsumption architecture emerged in the 1990s due to the need for more intelligent robot controllers that model robot behaviours that are reactive to environment changes and sensor measurement errors. The subsumption architecture was developed in the study [28]. The implementation consists of networks made of reactive controllers. A node or state often contains sensor conditions that determine whether to stay in or leave that state. It is possible to go from one state (the transmitter) to another (the receiver) through arcs containing messages that are the output of the recently left state. Sometimes, these arcs represent physical wiring, and the messages are the signals sent to the robot’s motor. These augmented finite state machines (AFSM) have internal clocks that control the time it takes to traverse an arc; the word augmented refers to such internal clocks. Figure 1. Below shows a model of a robot that uses (AFSM) and explains its working mechanism. Two or more simple networks representing simple behaviours can be connected one to another to form more complex networks and thus represent more complex behaviours, as mentioned in the study [29]. This process is called layering.

1.png

Figure 1. A hexapod robot called Genghis (left) was created by Brooks [30] using 57 augmented finite states machines (AFSM). An AFSM for controlling only one leg is illustrated on the right side. The AFSM receives feedback from a proximity sensor in order to see if the leg is stuck or not and to lift it higher next time

Machine learning is a powerful AI technique that has revolutionized the field of mobile robotics. Machine learning involves training a model on a dataset to make predictions or decisions. In mobile robot control, machine learning has been successfully applied in various tasks such as object detection, image recognition, path planning, and more. There are two main types of machine learning: supervised learning and unsupervised learning. Supervised learning involves training a model on labeled data, where the input data is paired with the correct output. The model learns to predict outputs based on input data by identifying patterns within the labeled dataset. Supervised learning algorithms find applications in a multitude of mobile robot control tasks, spanning from object detection and image recognition to path planning. For example, a mobile robot can be trained to recognize different objects in its environment using supervised learning. On the other hand, unsupervised learning involves training a model on unlabeled data, where the input data is not paired with any output. The model learns to discern patterns and structures within the data without any explicit guidance. Unsupervised learning algorithms contribute significantly to mobile robot control, serving purposes such as clustering, anomaly detection, and feature extraction. For example, unsupervised learning can be used to identify anomalies in sensor data, which can be indicative of a malfunctioning component in the robot. machine learning is a powerful AI technique that has been successfully applied in various mobile robot control tasks. Supervised learning involves training a model on labeled data, while unsupervised learning involves training a model on unlabeled data. Both approaches have their unique advantages and applications in mobile robot control, making them valuable tools for creating intelligent and adaptive mobile robots.

Machine learning (ML) is an expanded group of algorithms that can utilize datasets to determine patterns, deepen understanding and make determinations [32]. It gives the tools needed to analyze structured or unstructured data, draw conclusions, and make predictions [33]. The goal of machine learning is to investigate the concepts, algorithms, and theories used in statistics, computer science, and optimization [34]. According to the study [35], a robot can be seen as a system that interacts with the world and uses learning as a method of improving its performance. Machine learning plays a pivotal role in shaping the landscape of robotics, but it is not without its set of challenges. To fully harness its potential, machine learning in robotics must confront several key obstacles: Firstly, it must acquire the capability to learn representations of the world that align with the specific objectives programmed into the robot. These representations serve as the foundation for the robot to comprehend and achieve its intended purpose. Secondly, machine learning should encompass the calibration of sensors and actuators [36-42]. This calibration process enhances the precision and accuracy of these critical components, enabling the robot to interact more effectively with its environment. Lastly, it's crucial for machine learning to address the nuances of a robot's behavior in relation to its perceived environment and internal state. Understanding and managing the intricate interplay between these aspects are paramount for effective learning and decision-making. Within the realm of mobile robot control, several types of machine learning algorithms come into play:

Supervised Learning stands out as a prominent approach, involving the training of machine learning models on labeled data. Through this method, models acquire the ability to predict outputs based on input data by identifying patterns within the labeled dataset. Supervised learning algorithms find applications in a multitude of mobile robot control tasks, spanning from object detection and image recognition to path planning. On the other hand, Unsupervised Learning takes a different route, relying on unlabelled data for training. In this paradigm, machine learning models autonomously discern patterns and structures within the data without any explicit guidance. Unsupervised learning algorithms contribute significantly to mobile robot control, serving purposes such as clustering, anomaly detection, and feature extraction. These techniques further enrich the repertoire of machine learning tools at the disposal of robotics researchers and practitioners.

One of the critical challenges in mobile robotics is robot localization, which involves determining the position and orientation of a robot in its environment. Robot localization is essential for enabling mobile robots to navigate autonomously and perform tasks accurately. Several techniques have been developed to address the problem of robot localization, including Bayes Filter, Kalman Filter (KF), Extended Kalman Filter (EKF), and Particle Filter. Bayes Filter is a probabilistic approach to robot localization that involves estimating the probability distribution of the robot's position and orientation based on sensor measurements. Bayes Filter has been successfully applied in various mobile robot control tasks, such as simultaneous localization and mapping (SLAM). For example, a mobile robot can use Bayes Filter to estimate its position and orientation based on sensor measurements such as odometry and laser range finders. Kalman Filter (KF) is another popular technique for robot localization that involves estimating the state of a system based on noisy sensor measurements. KF is widely used in mobile robot control tasks such as tracking and navigation. For example, a mobile robot can use KF to estimate its position and velocity based on sensor measurements such as GPS and inertial sensors. Extended Kalman Filter (EKF) is an extension of KF that can handle nonlinear systems. EKF is widely used in mobile robot control tasks such as localization and mapping. For example, a mobile robot can use EKF to estimate its position and orientation based on sensor measurements such as visual odometry and laser range finders. Particle Filter is a probabilistic approach to robot localization that involves estimating the probability distribution of the robot's position and orientation based on a set of particles. Particle Filter has been successfully applied in various mobile robot control tasks, such as SLAM and tracking. For example, a mobile robot can use Particle Filter to estimate its position and orientation based on sensor measurements such as visual odometry and laser range finders. Robot localization is a critical challenge in mobile robotics, and several techniques have been developed to address this problem. Bayes Filter, Kalman Filter (KF), Extended Kalman Filter (EKF), and Particle Filter are some of the popular techniques used in mobile robot control tasks such as localization, mapping, and tracking. These techniques have enabled mobile robots to navigate autonomously and perform tasks accurately, making them valuable assets in various industries.

Localization, a critical process in robotics, revolves around determining a robot's precise position within its surroundings. When exploring the diverse landscape of localization techniques, several key comparisons come to light: Firstly, there's the distinction between Local and Global Localization. In scenarios where a robot lacks any prior information about its initial position, it must rely on sensor observations to infer its pose. This approach extends to challenging situations like the kidnapped robot problem, wherein a robot aware of its environment suddenly finds itself relocated to an unfamiliar location [78]. Another crucial consideration is whether localization occurs in Static or Dynamic Environments. The localization process must adapt depending on whether the robot operates in a stable, unchanging setting or contends with a dynamic environment characterized by moving entities [79]. Dynamic environments introduce added complexity as the constantly changing landscape necessitates the incorporation of dynamic elements into the environment state. Furthermore, the dimension of Single-robot versus Multi-robot Localization warrants attention. In the former, the focus revolves around a solitary robot navigating the environment, whereas the latter scenario entails deploying a group of robots within the same surroundings. Interestingly, the insights gleaned from single-robot localization can often be applied individually to each robot within a multi-robot setting, facilitating collaborative localization efforts [80, 81]. These distinctions underscore the intricate web of choices and considerations inherent in the realm of robotics localization. Sensors and actuators take part in determining the robot’s localization, but both are subjected to noise, and thus the problem of localization becomes difficult [82]. Another problem with sensors is that they usually do not provide enough information content to determine the position of the robot directly, and therefore, the robot might be in an ambiguous location. This problem is known as sensor aliasing Siegwart: intro-autonomous-robots and, along with sensors and actuators noise turns the localization problem into a difficult task. Here are some known techniques to deal with robot localization are mentioned:

(1) Bayes Filter: As claimed by Fox et al. [83], Bayes Filter is a statistical estimator of a dynamic system’s state based on noisy observations. In other words, it is an algorithm that calculates the belief’s distribution based on measurements and control data [84]. This is generally done in two steps: the prediction and the correction step, and thus each time a robot receives the sensor’s measurement data, the robot controller software needs to compute the posterior densities but notice that such a task is computationally heavy. Consequently, its time complexity grows exponentially because the amount of sensor’s measurements increases over time. A solution to this problem is to assume that such a dynamic system has the Markovian Property. That is, the future state x_t+1 depends on the present state x_tbecause of the assumption that x_tcarries all the needed information. Bayes filter algorithm is recursive. It requires the previous posterior distribution to calculate the current posterior distribution.

(2) Kalman Filter (KF): is a Gaussian technique invented in 1950 by Kalman [85] and was first implemented by NASA in the Apollo program to estimate the trajectory of the space capsule in real time [86]. It takes noisy data and takes the noise out to get information with less uncertainty [87]. It works with continuous states and represents the beliefs by the first and second moments [88] from multivariate normal distributions. Laskar et al. [89] proposed another version of KF called Maximum Correntropy Kalman Filter (MCKF), an effective and robust algorithm for non-Gaussian signals and heavy-tailed impulsive noise. Instead of using the Minimum Mean Square Error (MMSE) as KF, its optimality criterion is the Maximum Correntropy Criterion (MCC).

(3) Extended Kalman Filter (EKF): does not assume any linearity as KF does, and therefore, the next state and the measurement probabilities are nonlinear functions, as mentioned in the study [89]. Some variants of this method include the Unscented Kalman Filter, where the state distribution is approximated by a Gaussian random variable as in EKF, but now a minimal set of sample points is cautiously chosen for representing the state distribution [90]. Another improvement of EKF is the Invariant Extended Kalman Filter (IEKF) used for continuous-time systems with discrete observations [91].

(4) Particle Filter: Early in the nineties, the particle filter gained popularity and has been utilized to resolve estimate problems [92]. It acts as an approximate numerical solution to nonlinear Bayesian filtering problems [93], where the posterior is approximated by a set of M samples called particles where M is usually a large number (e.g., 2000). Under the context of localization, it is also known as Monte Carlo localization (MCL) [94]. Each particle has associated a weight, also called the importance factor, that represents the contribution to the overall estimate of the posterior, as mentioned in the study [95]. Thus, it shows the belief at time t approximated by a set of particles and weights of M particles [96-98].

Probabilistic robotics is a promising field that has the potential to revolutionize the field of mobile robotics. Probabilistic robotics employs probabilistic techniques and algorithms to allow robots to make better decisions in the face of uncertainty. The key idea behind probabilistic robotics is to use probability theory to model uncertainty and to develop algorithms that can reason about this uncertainty. This is particularly important in real-world environments, where sensors can be noisy, and the robot's state and environment can be uncertain. Probabilistic robotics is based on three main components: perception, control, and state estimation. Perception refers to the process of sensing the environment and extracting information from the sensors. Control refers to the process of selecting actions that achieve some objective. State estimation refers to the process of estimating the state of the robot and its environment. Probabilistic techniques and algorithms are applied in perception, control, and state estimation in the context of mobile robotics. For example, in perception, probabilistic techniques such as Bayesian inference and Markov models are used to estimate the probability distribution of the robot's environment based on sensor measurements. In control, probabilistic techniques such as optimal control and reinforcement learning are used to select actions that maximize some objective function. In state estimation, probabilistic techniques such as Kalman filters and particle filters are used to estimate the probability distribution of the robot's state based on sensor measurements.

Probabilistic robotics has been utilized in various mobile robot control tasks, such as robot localization and obstacle avoidance. For example, in robot localization, probabilistic techniques such as particle filters and Kalman filters have been used to estimate the robot's position and orientation based on sensor measurements. In obstacle avoidance, probabilistic techniques such as fuzzy logic and Bayesian networks have been used to detect and avoid obstacles in the robot's environment. probabilistic robotics is a promising field that has the potential to revolutionize the field of mobile robotics. Probabilistic techniques and algorithms are applied in perception, control, and state estimation in the context of mobile robotics, enabling robots to make better decisions in the face of uncertainty. Probabilistic robotics has been successfully applied in various mobile robot control tasks, such as robot localization and obstacle avoidance, making it a valuable tool for creating intelligent and adaptive mobile robots.

This section explains the critical and relevant concepts from probabilistic robotics applied to the current research. Probabilistic robotics is a promising field that has the potential to revolutionize the field of robotics. Probabilistic robotics has several applications, including autonomous driving and mobile robotics. In mobile robotics, probabilistic robotics is used to localize the robot in the environment and to plan paths that avoid obstacles. It employs probabilistic techniques and algorithms to allow robots to make better decisions in the face of uncertainty, as mentioned in the study [104]. Probabilistic robotics is based on three main components:

(1) Perception: Perception refers to the process of sensing the environment and extracting information from the sensors [105].

(2) Control: Control refers to the process of selecting actions that achieve some objective.

(3) State estimation: State estimation refers to the process of estimating the state of the robot and its environment.

The key idea behind probabilistic robotics is to use probability theory to model uncertainty and to develop algorithms that can reason about this uncertainty. In traditional robotics, uncertainty is often treated as a nuisance that needs to be eliminated. In probabilistic robotics, uncertainty is explicitly modelled and used to improve decision-making, as mentioned in the study [100]. This is particularly important in real-world environments, where sensors can be noisy, and environments can be dynamic and unpredictable. Probabilistic robotics provides a robust framework for dealing with the challenges of uncertainty in real-world robotic applications. Robots can make more informed decisions by modelling and accounting for uncertainty, improving performance and reliability.

Differentiable programming (DP) is a programming paradigm that enables the automatic differentiation of algorithms. DP is a combination of the ideas of automatic differentiation and functional programming. DP has demonstrated that it is more than just a compilation of machine learning algorithms. DP can be applied to various types of robots, including mobile robots, manipulators, and humanoid robots. DP can be used for tasks such as trajectory planning, motion control, and reinforcement learning algorithms. DP works by allowing the user to define a function that maps inputs to outputs. The function can be a complex algorithm that involves multiple operations, such as matrix multiplication, convolution, and activation functions. DP then automatically computes the gradient of the function with respect to its inputs using the chain rule of calculus. The gradient can be used to optimize the function using gradient-based methods such as stochastic gradient descent (SGD). DP has several advantages over traditional programming approaches for controlling robots. DP enables the efficient optimization of control policies using gradient-based methods, which can lead to faster convergence and better performance. DP also allows the integration of sensor feedback into the control policy, which can improve the robot's ability to adapt to changing environments. DP is a promising technology that has the potential to revolutionize the field of robotics.

DP has been successfully applied in various mobile robot control tasks, such as trajectory planning, motion control, and reinforcement learning algorithms. For example, in trajectory planning, DP can be used to optimize trajectories by differentiating through the trajectory generation algorithm. The objective function can be defined as a function of the robot's state and the control inputs. DP can be used to compute the gradient of the objective function with respect to the control inputs, which can be used to update the control policy. In motion control, DP can be used to optimize the control policy by differentiating through the dynamics of the robot. DP can be used to compute the gradient of the control policy with respect to the robot's state and the control inputs, which can be used to update the control policy. In reinforcement learning, DP can be used to optimize the Reinforcement Learning algorithm by differentiating through the policy evaluation and improvement steps. DP can be used to compute the gradient of the policy with respect to the parameters of the policy, which can be used to update the policy. DP is a promising technology that enables the automatic differentiation of algorithms.

In order to understand what differentiable programming (DP). The study [35] is restated here. DP is a programming paradigm that enables the automatic differentiation of algorithms. DP is a combination of the ideas of automatic differentiation and functional programming. It has demonstrated that it is more than just a compilation of machine learning algorithms [109]. DP can be applied to various types of robots [110], including mobile robots, manipulators, and humanoid robots. DP can be used for tasks. In this section, we will discuss some of the applications of DP in controlling robots.

(1) Trajectory Planning: Trajectory planning is the process of generating a sequence of robot poses that satisfy some objective. It is an essential problem facing mobile robotics [111] and is computationally extremely costly [112]. DP can be used to optimize trajectories by differentiating through the trajectory generation algorithm. The objective function can be defined as a function of the robot's state and the control inputs. DP can be used to compute the gradient of the objective function with respect to the control inputs, which can be used to update the control policy.

(2) Motion Control: Motion control is the process of controlling the robot's motion to achieve some objective [113]. DP can be used to optimize the control policy by differentiating through the dynamics of the robot.

(3) Reinforcement Learning: DP can be used to optimize the Reinforcement Learning algorithm by differentiating through the policy evaluation and improvement steps.

DP has several advantages over traditional programming approaches for controlling robots. DP enables the efficient optimization of control policies using gradient-based methods [114], which can lead to faster convergence and better performance. DP also allows the integration of sensor feedback into the control policy, which can improve the robot's ability to adapt to changing environments. DP is a promising technology that has the potential to revolutionize the field of robotics. However, there are some limitations of DP that need to be addressed. DP requires the algorithm to be differentiable, which can limit the complexity of the algorithm that can be used. DP also requires the computation of the Jacobian matrix, which can be computationally expensive for large systems.