[Paper Review] Data-driven Methods Applied to Soft Robot Modeling and Control: A Review

Zixi Chen, Federico Renda+7 more|arXiv (Cornell University)|May 20, 2023

Cardiovascular Health and Disease PreventionMedicine3 citations

TL;DR

This review synthesizes data-driven modeling and control methods for soft robots, categorizing them into five types—Jacobian, analytical, statistical, neural networks, and reinforcement learning—comparing their data needs, control performance, and suitability for tasks. It identifies that hybrid offline-online learning approaches will likely dominate future soft robot control due to trade-offs between model complexity, computation speed, and adaptability to robot aging and variability.

ABSTRACT

Soft robots show compliance and have infinite degrees of freedom. Thanks to these properties, such robots can be leveraged for surgery, rehabilitation, biomimetics, unstructured environment exploring, and industrial grippers. In this case, they attract scholars from a variety of areas. However, nonlinearity and hysteresis effects also bring a burden to robot modeling. Moreover, following their flexibility and adaptation, soft robot control is more challenging than rigid robot control. In order to model and control soft robots, a large number of data-driven methods are utilized in pairs or separately. This review first briefly introduces two foundations for data-driven approaches, which are physical models and the Jacobian matrix, then summarizes three kinds of data-driven approaches, which are statistical method, neural network, and reinforcement learning. This review compares the modeling and controller features, e.g., model dynamics, data requirement, and target task, within and among these categories. Finally, we summarize the features of each method. A discussion about the advantages and limitations of the existing modeling and control approaches is presented, and we forecast the future of data-driven approaches in soft robots. A website (https://sites.google.com/view/23zcb) is built for this review and will be updated frequently.

Motivation & Objective

• To systematically categorize and compare data-driven models used in soft robot modeling and control.
• To identify the strengths, limitations, and trade-offs of five major data model categories: Jacobian, analytical, statistical, neural networks, and reinforcement learning.
• To address the challenges of model transferability, robot aging, and data reliability in soft robotics.
• To propose a future direction favoring hybrid offline-trained and online-learning controllers for improved adaptability and performance.
• To highlight the safety and interpretability gap in applying deep learning and RL in medical soft robotics.

Proposed method

• Classifying existing soft robot modeling and control approaches into five data model categories based on their mathematical and computational foundations.
• Analyzing each model type using criteria such as data requirements, control frequency, dynamic capability, and online update support.
• Surveying recent literature (post-2019) to map model applications across robot designs, sensors, and tasks.
• Evaluating model performance through comparative analysis of control accuracy, computational load, and feasibility in real-time applications.
• Proposing a hybrid control framework combining offline-trained models with online adaptation to address model drift and robot variability.
• Highlighting the role of simulation and physical simulators in training RL agents due to the infeasibility of in vivo exploration.

Experimental results

Research questions

• RQ1 How do different data-driven models (Jacobian, analytical, statistical, neural networks, RL) compare in terms of data requirements, control frequency, and modeling accuracy for soft robots?
• RQ2 What are the key limitations of current data-driven models in soft robot control, particularly regarding model transferability and long-term reliability?
• RQ3 Why is interpretability a major barrier for deploying neural networks and reinforcement learning in medical soft robotics?
• RQ4 How can online learning compensate for the performance degradation of soft robots due to material aging and manufacturing variability?
• RQ5 What is the optimal balance between model complexity and computational efficiency for real-time soft robot control?

Key findings

• Neural networks and reinforcement learning achieve high performance on nonlinear and complex deformations but require large datasets and are unsuitable for online training due to computational complexity.
• Jacobian and statistical models (e.g., GPR, KF) support online learning and higher control frequencies, making them more practical for real-time applications despite lower modeling accuracy.
• Analytical models such as FEM and state-space representations offer detailed physical descriptions but are not suitable for online adaptation or real-time control.
• The combination of offline-trained models and online-learning components is predicted to become the dominant control strategy due to improved robustness against robot aging and inter-robot variability.
• Reinforcement learning cannot be trained in vivo and must rely on simulation, limiting its clinical deployment despite high-level task capabilities.
• Interpretability and safety concerns remain major barriers to deploying deep learning and RL in medical applications, necessitating interdisciplinary collaboration.