Deep Learning to model Soft Robot Arms and Hands

With the increased interest in the use of soft materials for the creation of highly dexterous robots, soft material robotics has established itself as an important research topic within robotics. Some roboticists predict that soft robotic technologies will play a key role in many areas, such as safe human-robot interaction, minimally invasive surgery, disaster scenarios, underwater and planetary exploration as well as grasping and manipulation of fragile objects, as for example, in food-related industries.

Although soft robots open up new avenues of application areas especially because of their inherent compliance and capability to safely interact with their environment, controlling their movements to conduct useful tasks is challenging. While it is straightforward to create kinematic models for the more traditional rigid-link robot arms and hands, most analytical models for soft robots are not capable of accurately dealing with the nonlinearities of robot arms made from soft materials and their difficult-to-predict motion behaviour especially when exposed to external forces imposed by interactions with the environment.

This project aims at exploring machine learning techniques, especially deep learning, to generate data-driven models for soft robot arms and hands. Motion data will be acquired from real and simulated soft robot arms and hands conducting a range of movements whilst conducting relevant tasks, such as picking and handling of fragile objects including fruits and vegetables.

Instead of learning neural networks from scratch, this project will make good use of analytical soft robot models (e.g., soft robot kinematic models) and improve on those through the learning approach. Since only aspects that are difficult to model will be learnt, the proposed approach is expected to get more robust training outcomes and require less training. The most important outcomes from the project will be robust kinematic representations and motion primitives for soft robot arms and hands interacting with the physical environment.