Advanced navigation systems are increasingly becoming crucial for robots due to their growing sophistication and usage in different applications, such as search and rescue operations, manufacturing, transportation, disaster relief, and defense. Artificial intelligence (AI) technologies can transform the way robots navigate in their environments, making them more adaptable, efficient, and capable of handling complex tasks. This article discusses the importance of AI in robot navigation and the key AI techniques used in this field.
Image credit: Generated using DALL.E.3
Importance of AI
Conventional robot navigation systems primarily depend on fixed sensors and pre-programmed instructions to navigate their environments. Although these robots can effectively navigate in controlled settings, they face significant navigation challenges in unpredictable or dynamic environments.
Thus, the robot can get stuck or fail to efficiently execute its task in such scenarios and require human intervention. AI-powered robot navigation systems can effectively address these issues by enabling robots to learn from their experiences and adapt to unknown situations.
Specifically, AI-powered navigation systems can enable robots to analyze and process substantial amounts of data in real time by incorporating advanced sensors and machine learning (ML) algorithms, which allow robots to make informed navigation-related decisions in different environments.
For instance, deep learning (DL) algorithms can be utilized to improve robot navigation. A DL algorithm can be trained using large datasets of sensor data/images to enable a robot to recognize several types of environmental features, such as terrain, and effectively navigate them.
Other AI technologies, such as reinforcement learning (RL), can be utilized to improve robot navigation. A robot can continually adjust its actions based on the feedback received by it using RL to learn to more effectively navigate its environment over time.
Moreover, computer vision technologies can be integrated with advanced sensors for better robot navigation. AI algorithms can develop a more detailed and accurate map of a robot’s environment by combining data obtained from different sensors, such as ultrasonic sensors, lidar, and cameras, to enable the robot to better comprehend its surroundings and make decisions about navigation.
AI can also enable communication and collaboration between robots to allow them to work more effectively in complex environments. For instance, a team of robots can utilize AI algorithms to share information about their environment to better coordinate their movements and prevent collisions. Thus, the use of AI technologies can facilitate the development of sophisticated robot navigation systems, which can positively impact several industries and applications.
AI Techniques in Robot Navigation
Several AI techniques can be used in robot navigation, including particle swarm optimization (PSO), ant colony optimization (ACO), genetic algorithm (GA), neural networks (NNs), fuzzy logic system, and deep reinforcement learning (DRL).
Fuzzy Logic Technique: Eight rule-based fuzzy controllers can be used for path following and obstacle avoidance for mobile robots, while gradient method-based Takagi Sugeon fuzzy controllers can tune various membership function parameters to acquire the optimal result for robot navigation.
Similarly, the Khepera simulator with fuzzy logic-based agents can be employed to control robots. The behavior of every agent, including sensor value, robot position, and heading angle, can be controlled by defining the sets of fuzzy rules. A memory system can be included to further increase the system's efficiency by enabling the robot to identify alternative routes when it gets trapped.
A fuzzy logic controller can be used for path following depending on the orientation and position errors. The control of two wheels independently using the controller can provide longitudinal and lateral control of the robot. Fuzzy logic controllers can also be utilized for sensor-based mobile robot navigation in indoor environments. Fuzzy controllers can be optimized by combining the RL and GA methods. Fuzzy logic with visual landmark recognition can be used for obstacle avoidance. The path following and control problem of autonomous mobile robots can be solved using an ultrasonic range finder by combining GA and fuzzy controller.
NNs: NNs can be used to solve several robot navigation problems, including defining schedules and identifying the shortest route for traveling. For instance, multilayer feed-forward artificial NN can be combined with the Q reinforcement method for effective path planning.
Similarly, a multilayer NN controller and proportional integral derivative (PID) can be utilized to design an Arduino microcontroller-based direct current (DC) motor for controlling speed in robots. NN architecture can also be used for designing an automatic steering controller for an autonomous mobile robot, and to develop a collision-free path in a dynamic environment.
A biologically inspired NN can be employed to develop a wall following robot, while a hybrid NN can be used for efficient robot navigation. Goal-seeking and obstacle-avoidance behaviors can be realized in robots using NN.
The trajectory tracking problem in robot navigation can be solved using the adaptive NN PID controller, while a combination of the first-order Sugeno fuzzy inference model and adaptive neuro-fuzzy inference system (ANFIS) can be utilized for coordinating several robots and path planning.
Two different NN controllers can be employed for path following and controlling robots. Additionally, Hopfield NN can be used for path planning and obstacle avoidance in complex environments. Multilayered NN and recurrent neural network (RNN) can be utilized for designing intelligent navigation systems for mobile robots and solving path following and localization problems, respectively. The RNN assists the robot in autonomously navigating an unknown environment.
Moreover, a type-2 fuzzy neural network (IT2FNN) can be employed to effectively address the obstacle avoidance and orientation stabilization of wheeled robots. IT2FNN possesses three layers, including the output, hidden, and input layer, and four inputs. Angular and linear velocities of the robot are the outputs of the robot. Dynamic nonholonomic robots can be controlled using NN.
GA: GAs can be used to solve path planning problems in both dynamic and static environments. The navigation path length of robots in a cluttered space can be optimized using the Petri-GA technique.
A fuzzy controller combined with GAs can be used for the guidance of robots in a static and dynamic environment and for optimizing the navigation path length. GA can be utilized to select the most suitable membership function parameters from a fuzzy inference system to control a robot’s steering angle in a partially unknown environment.
The optimal path for a robot can be identified using GA and fuzzy logic, while effective path planning of several robots can be achieved using an improved GA, which can guide robots efficiently from the origin to the destination without any collision.
Motion control can be realized by implementing a genetic-fuzzy controller (GA-FLC) for tuning and optimizing the Gaussian membership function parameters. Additionally, multiple objective genetic algorithm (MOGA) and single fitness-based GA can be employed for path optimization of the robot and avoiding navigation problems in the dynamic environment, respectively.
PSO: The motion planning problem of a robot can be solved using multi-objective PSO, while the velocities of the left motor and right motor of the differential drive robot can be determined using a PSO-based optimal fuzzy controller.
The parallel met heuristic PSO (PPSO) algorithm can be used to address the global path-planning problem of robots. Moreover, an evolutionary-group-based PSO (EGPSO) for automatic learning of fuzzy systems can be utilized for wall following control and robot navigation.
DRL: Uncertainty-aware RL, double deep Q network (DDQN), asynchronous deep deterministic policy gradient (DDPG), fast recurrent DPG, and successor feature RL can be utilized for local obstacle avoidance.
Long short-term memory (LSTM) + DRL, asynchronous advantage actor-critic (A3C) + LSTM, and LSTM + proximal policy optimization (PPO) can be utilized for indoor navigation. Additionally, PPO, parallel DDPG, parallel PPO, and collision avoidance with DRL can be employed for multi-robot navigation.
Recent Studies
Autonomous robotic navigation to a specified point is primarily achieved using a layered stack of local motion planning and global path planning modules that generate obstacle-free and feasible trajectories. Although these modules can be modified to fulfill user preferences and task-specific constraints, existing modification procedures need the substantial efforts of an expert roboticist with great technical expertise.
In a study published in the journal Frontiers in Robotics and AI, researchers investigated the feasibility of simplifying this process by inserting an ML module between the local motion planning and global path planning modules of an off-the-shelf navigation stack.
This model was trained with human demonstrations of the preferred navigation behavior using a behavioral cloning-based training procedure, which allowed an intuitive navigation policy modification even by non-technical users to meet the task-specific constraints. Results demonstrated that the proposed approach can adapt a robot’s navigation behavior successfully, leading to a trajectory similarity to the demonstrator and a higher navigation success rate compared to baselines.
References and Further Reading
Warnell, G., Stump, E., Kochersberger, K., Waytowich, N. R. (2021). Improving Autonomous Robotic Navigation Using Imitation Learning. Frontiers in Robotics and AI, 8, 627730. https://doi.org/10.3389/frobt.2021.627730
Zhu, K., Zhang, T. (2021). Deep reinforcement learning based mobile robot navigation: A review. Tsinghua Science and Technology, 26, 5, 674-691. https://doi.org/10.26599/TST.2021.9010012.
Gul, F., Rahiman, W., Alhady, S. S. N. (2019). A comprehensive study for robot navigation techniques. Cogent Engineering. https://doi.org/10.1080/23311916.2019.1632046
Frąckiewicz, M. (2023). AI in Robot Navigation. [Online] Available at https://ts2.space/en/ai-in-robot-navigation/ (Accessed on 13 November 2023)