AI-based Design Assistance System for Soft Robotics – Integrating Motion and Control in the Design Process

Robotic systems made entirely of soft material, without explicit joints or backbone structure, are compliant, adaptable and can deform severely, which makes them prone for wearable devices, domestic robots, medicine, assistance, legged robots and grasping. However, due to the highly elastic structure, primarily made of visco-elastic material with nonlinear characteristics, and their continuous deformation, the design faces new challenges in modeling and predicting the behavior of these systems.

The aim of this project is to develop a design assistance system based on profound knowledge of the smallest design entity of a soft robotic system – in our case this is the pressurizable silicone cylinders of a universal soft bending actuator – to define the optimal design within the multi-parameter space in geometry, actuation and control.

The design assistance system will be based on underlying models and analysis of the forward kinematics and dynamics with help of a hierarchical approach of physical models and machine learning algorithms. By just defining the desired path that the universal bending actuator shall fulfill and the loads acting on the structure, the design assistance system shall determine the optimal dimension of all components, as well as the number of actuators connected in series and their actuation. Additional information, as the stress distribution in the bending actuator, the static and dynamic deformation and orientation, and the workspace shall be given as an output. Additional experience of the designer or applicator can be additionally integrated into the design process.

Inspired by numerous developed soft robots, we conducted an extensive study on various design aspects to develop a universal soft bending actuator that is easy to fabricate, simple to model, and broad in applicability.

Important design aspects studied are:

• the bending principle (bending air chambers vs. stretching/contracting air chambers),
• the dimensions: cross section, number and connection of the pressurisable air chambers,
• the number and reinforcement of air chambers.

We found the design of three parallel aligned, cylindrical air chambers with ring-reinforcement, tied together with spacers, and closed by end caps is optimal for combined deformations (bending and stretching) and at the same time easy to manufacture. An advantage of our modular design is that each part can be easily adapted to realize specific functionality by either change the dimension or the material. Moreover, the modeling of this soft bending actuator is straightforward and numerical solutions exist for the characteristic behavior of a single pressurizable air chamber, from which the behavior of the entire module can be derived.

The adaptation of the basic design (dimensions, material, reinforcement) shall be realized with the design assistance system and the artificial neural network, which can select the best suited design from the large database obtained from numerical simulations.

As part of the first phase, a design assistant framework has been established. This framework is capable of optimizing the cross-sectional structure of soft pneumatic bending actuators such that a given target workspace can be fitted.

The design space shifted slightly from the approach initially envisioned to allow for more variation in possible designs. Still, this new design space considers the cross-sectional design of the universal bending actuator described above while, for example, being able to assign multiple inner air chambers to the same pressure supply.

In addition, this design assistant system is capable of considering manufacturability criteria. These criteria include, for example, lower bounds for the wall thickness and a preferable low number of individual chambers. This helps in reducing errors in the silicone molding process. The current state of the design assistant system with additional features planned for the second phase is pictured below.

In the second phase, we intend on expanding the existing design assistant system. While a prominent advantage of soft robots is their softness when interacting with objects or living entities, this interaction is currently not considered. For this, we plan to implement a contact model capable of presenting interaction with objects, surfaces, or viscous mediums. A further objective in the second phase will be the implementation of path planning and control as part of the assistant system, forming a Design and Control Assistant System for Soft Robots.

When optimizing for a design in the first phase, it became obvious that more sophisticated material models and more acurate material parameters are needed to correctly describe the behavior of the soft robot. This motivated another part of the second phase: investigating visco-elastoplastic material models and additionally making the existing design assistant more stable regarding manufacturing uncertainties. To derive parameters to fit this material model, a large magnitude of cyclic tensile tests over a period of 30 months with varying storage conditions are conducted.

• Numerical work
  • FEM model
  •  Beam model
• Experimental work
  • Exchange of test equipment
  •  Providing prototypes
• Machine learning methods
  •  Exchange of NN
  •  Training our NNs with other data

• Data generation
  • Including results from additional models
  • Adding further aspects
• Machine learning algorithms
  • Exchange of experiences
  • Collaboration on NN
• Design assistance system
  •  Exchange of experiences