How Scientists Are Teaching Soft Robots to Move Accurately

:::info Authors:

(1) Jorge Francisco Garcia-Samartın, Centro de Automatica y Robotica (UPM-CSIC), Universidad Politecnica de Madrid — Consejo Superior de Investigaciones Cientıficas, Jose Gutierrez Abascal 2, 28006 Madrid, Spain ([email protected]);

(2) Adrian Rieker, Centro de Automatica y Robotica (UPM-CSIC), Universidad Politecnica de Madrid — Consejo Superior de Investigaciones Cientıficas, Jose Gutierrez Abascal 2, 28006 Madrid, Spain;

(3) Antonio Barrientos, Centro de Automatica y Robotica (UPM-CSIC), Universidad Politecnica de Madrid — Consejo Superior de Investigaciones Cientıficas, Jose Gutierrez Abascal 2, 28006 Madrid, Spain.

:::

Abstract and 1 Introduction

2 Related Works

2.1 Pneumatic Actuation

2.2 Pneumatic Arms

2.3 Control of Soft Robots

3 PAUL: Design and Manufacturing

3.1 Robot Design

3.2 Material Selection

3.3 Manufacturing

3.4 Actuation Bank

4 Data Acquisition and Open-Loop Control

4.1 Hardware Setup

4.2 Vision Capture System

4.3 Dataset Generation: Table-Based Models

4.4 Open-Loop Control

5 Results

5.1 Final PAUL version

5.2 Workspace Analysis

5.3 Performance of the Table-Based Models

5.4 Bending Experiments

5.5 Weight Carrying Experiments

6 Conclusions

Funding Information

A. Conducted Experiments and References

If in the design of pneumatic soft robots there is a wide variety of possibilities due to the still immature state of these robots, in control the possibilities increase even more. Indeed, controlling soft robots is still an open challenge [51] and there are many possible solutions, even if they do not yet achieve similar precision to those existing in rigid robotics.

\ Both model-based and data-driven control methods have been tested. Although some authors suggest that, unlike what has happened in other areas, where it started with model-based controllers and has evolved to the use of Machine Learning techniques, here the opposite is happening [52], for the moment both techniques coexist and it does not seem that, in the short term, either philosophy is going to impose itself.

\ Model-based techniques include the use of PCC on the one hand and FEM on the other. The former has been used successfully in pneumatic soft robots [50, 53]. Its main drawback, however, is that it is based on very strict premises –deformation with constant curvature and absence of gravity– which often make its application unfeasible.

\ The use of FEM, on the other hand, can achieve very good results. As pointed out by [52], although numerical methods are theoretically less accurate than analytical methods, when modelling soft structures such as these, analytical methods are not capable of working with such complex –or sometimes simply unknown– shapes, constitutive laws and boundary conditions.

\ Its use has been widely implemented in soft robots [12, 54]. In the case of pneumatic soft robots, the works of Ding [55] –which is, however, much smaller in size than PAUL–, or Cangan [56], which uses previously presented SoPrA arm, stands out. The problem they present, however, is the high computational cost required, which often makes their closed chain control unfeasible, unless reduced order models are used [57]. Furthermore, setting the different elastic and geometric parameters of the materials is not an automatic task due to how difficult it can be to characterise them.

\ Thus, different Machine Learning (ML) techniques are usually used to solve the modelling and control problem. Many techniques have been used, from Feedforward Neural Networks (FFNN) [45] to more complex network architectures [25, 58] in addition to the use of Reinforcement Learning [59, 60]. A wide variety of input data can be used to generate a fairly accurate model: data-driven models have been tested to perform well with input from the real robot [61], data from a FEM model [12], a combination of both [62], and visual data [63].

\ No relationship has been observed between the complexity of the ML technique used and the results obtained. On the contrary, the philosophy in these cases is to always use the simplest tool possible –which usually also requires a smaller amount of data– that achieves the expected results. Although it is not common, controls based on polynomial adjustments [5] or minimisation of cost functions [64] have even been achieved. This is the philosophy that has been followed with PAUL. The objective, furthermore, in this first stage, is not to achieve very accurate control –the implementation of sensors in it is expected as future work– but to have a model that allows carrying out initial movement tests.

\

:::info This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license.

:::

\