This review summarizes the data models applied in soft robot modeling and control. The Jacobian model is an initial viable choice due to its simple structure and success in rigid robot control. Such a method models soft robots with strong assumptions, which are the linear relationships, but can obtain a high control frequency. Some scholars also try to build analytical models based on the physical structure of soft robots and from the view of control systems. These models provide a deep understanding of soft robot modeling, but complex models are unsuitable for control. Statistical models are utilized to extract motion features from a moderate amount of data and cope with properties like nonlinearity and hysteresis. They summarize the characteristic of data without the requirement of physical knowledge, and some models can be applied for modeling and control simultaneously. With the development of NN, it has been proven that NN is an effective tool for modeling and control since it contains various structures, some of which are able to process sequential data. However, in most cases, it requires a large amount of data and can only be trained offline, owing to its learning strategy, network structure, and parameter number. Recently, RL has been a popular choice for soft robot control, and it can achieve some high-level and complicated tasks without modeling. Following these benefits, RL requires a huge amount of data, a realistic exploring environment, and a long training time. The summary of these model categories is shown in Table 6.
It is apparent that sophisticated models require a larger amount of data while obtaining better modeling and control performance, but they also reduce the computation frequency. Compared with simple local linear models like the Jacobian model, analytical and statistical models consider the nonlinearity of soft robots and obtain wider feasible areas. Moreover, NN and RL can solve issues like the memory effect, but they require more training data and time, even an exploring environment. When using NN, it is nearly impossible to train it online due to its complex structure. However, statistical models can learn online and obtain relatively high control frequency thanks to the concise structures. Oversimplified models are only feasible for limited simple tasks, and laborious models attain good modeling and control performance. Considering both cost and performance, each model has its own pros and cons, and there is no optimal solution. These days, there is a trend that most works focus on complicated RL in soft robot control, and models with a considerable amount of parameters are also welcome in other areas like ChatGPT. Thanks to the development of hardware like GPU, such huge models are feasible now. However, large models lead to low frequency for control implementation, and one should consider balancing the model complexity and computation cost based on the modeling and control requirement.
Some challenges have yet to be addressed in soft robot modeling and control. In soft robot manufacture, it is challenging to make two same robots compared with the rigid counterparts. Also, a soft robot will age and show different properties at different times. Although the commonly used modeling and control approaches show good performance, such strategies may fail to transfer to the other robots or even the same robot after a period. From the view of data-driven methods, the collected data is unreliable, which has also been mentioned by the previous reviews. To address this issue, one may propose an offline-trained model based on an available robot and include an online-learning model to compensate for the gap between the robot producing the dataset and the robot employed for testing. Genearlly speaking, the combination of various controllers may be a good solution by take advantages of all of the controllers. The medical environment, as one of the most significant applications of soft robots, has a high standard for safety, but NN and RL lack interpretability and are challenging to apply in real surgery. Also, it is impossible for RL to explore in vivo, and RL agents can only be trained in simulation or physical simulators. The automatic medical soft robot control is still in its nascence from the aspect of safety and data requirements. Cooperation among robotics researchers, doctors, and related departments is required to solve this issue.