Abstract
Due to their adaptability, flexibility, and deformability, soft robots have been widely studied in many areas. On the other hand, soft robots have some challenges in modeling, design, and control when compared to rigid robots, since the inherent features of soft materials may create complicated behaviors owing to non-linearity and hysteresis. To address these constraints, recent research has utilized different machine learning algorithms and meta-heuristic optimization techniques. First and foremost, the study looked at current breakthroughs and applications in the field of soft robots. Studies in the field are grouped under main headings such as modelling, design, and control. Fundamental issues and developed solutions were analyzed in this manner. Machine learning and meta-heuristic optimization-oriented methods created for various applications are highlighted in particular. At the same time, it is emphasized how the problems in each of the modeling, design, and control areas impact each other.
TopIntroduction
As the use of robots grows, it may be necessary to redefine flexibility once more. While speed was one of the most essential factors for efficiency expected from robots confined to factories in the early stages of industrialization, flexibility could be considered as a margin of error. This approach may be appropriate for indoor applications. However, in outdoor applications, the opposite of this approach is mostly valid. In fact, since the ideal conditions in the external environment disappear, the adaptation of traditional robots becomes very difficult with their cumbersome structure. When we look at the living nature, the reason for this is better understood. Because one of the features that ensures the existence of life even under the most extreme conditions is flexibility. Gerringer et al. (2017), for example, found that a fish species (hadal belt snail fish) can survive in the Mariana Trench at a depth of about 8 km and a pressure of around 800 atm. The adaptability brought about by flexibility enables the living thing to accept even the most difficult conditions (very high pressure, etc.) as new conditions, not as a disturbance effect, and to continue its standard behaviors such as feeding and reproduction. The development of the idea of benefiting from flexibility and deformation, inspired by living things with such examples, has led to the emergence of soft robots.
Figure 1. Classification of robots by materials and degrees of freedom (Trivedi, 2008)
Soft robotic technologies have opened the door to a new era in robot design and development. However, these technological developments face new challenges and limitations due to the hyper-redundancy (Figure 1) and high degree of freedom caused by the flexible nature of the materials used. It also causes complex and unexpected behavior as a result of the nonlinear relationship between the input and output of the system.
The increasing complexity of soft robotic systems, which may be due to shape, material, actuation, and their complicated connection, renders conventional robot design methods inapplicable (Chen & Wang, 2020). The simulation and analytical tools required to effectively predict the complex mechanical behavior of soft robots are insufficient. In addition, there was a lack of effective optimization algorithms to automate the design process in the literature. To bridge this gap, optimization techniques have started to be implemented in the design of the soft robotics field.
On the other hand, the control challenges can be considered the final challenge to overcome before a soft robot can be utilized in the physical world. Although designs that use on-off logic, such as the soft robotic gripper, do not require control, this cannot be said for the other soft robot kinds. High-fidelity soft systems models have been effectively used to control soft systems, but this needs precise system identification and costly run-time calculation, hence restricting the applicability (Chin, Hellebrekers & Majidi, 2020). This high dimensionality is also a basic difficulty for soft sensing since many physical states might yield the same sensor readout once the higher dimensional data is converted. When traditional mathematical and statistical analytical models are deemed insufficient, data-driven machine learning (ML) control approaches have generated promising outcomes.
This chapter presents a comprehensive overview of ML-based studies and optimization methods in soft robotics, examines the current trends, and addresses the current limits of these algorithms. It also explains the application of ML algorithms and optimization techniques in soft robotics problems where the conventional methods continue to be ineffective. Furthermore, it analyses current ML-based studies in detail, covering learning types, tasks, and algorithms. The chapter also aims to provide an insight into researchers for their potential applications in this area. Finally, the advantages and some drawbacks of using ML techniques in soft robotics and future directions in this topic are provided in this chapter.
Key Terms in this Chapter
Clustering: Clustering is a technique for grouping a collection of items into clusters based on their similarity.
Classification: Classification is a commonly applied supervised learning method that assigns one of the predefined classes to new instances.
Machine Learning: Machine Learning is an area of artificial intelligence in which computers can self-learn based on past experiences.
Meta-Heuristic Optimization: Meta-heuristic optimization is the collection of operations and models that employ randomness to optimize the candidates and discover the optimum solution.
Soft Robotics: Soft robotics is a branch of robotics that focuses on technologies that mimic the physical features of live beings.
Regression: Regression is another supervised learning task that discovers correlations between data features and predicts continuous-valued target output based on the historical data.
Reinforcement Learning: Reinforcement learning aims to construct actions or strategies for learning anticipated behaviors by developing reward functions.
Association Rule Learning: Association rule learning, which is also a form of unsupervised learning based on rules, is a well-researched technique for uncovering interesting correlations and common patterns among items in the datasets.
Supervised Learning: The supervised learning technique aims to train an ML model using pre-labeled data.
Unsupervised Learning: Unsupervised learning technique learns patterns from unlabeled observation data with no previous knowledge of the output value.