Martin Stoelen, Plymouth University

Co-Exploring Actuator Antagonism and Bio-Inspired Control/Learning in a Printable Robot Arm

The human arm is capable of performing fast targeted movements with high precision, but is inherently ‘soft’ due to the muscles, tendons and other tissues of which it is composed. Robot arms are also becoming softer, to enable robustness when operating in real-world environments. But softness comes at a price, typically an increase in the complexity of the control required for a given task speed/accuracy requirement. This talk will introduce the GummiArm, a ‘soft’ robot arm that can almost entirely be printed on hobby-grade 3D printers. This enables rapid and iterative co-exploration of ‘brain’ and ‘body’, and provides a great platform for developing adaptive and bio-inspired behaviours. Viscoelastic actuator-tendon systems in an agonist-antagonist setup provide the arm with inherent damping, and stiffness that can be varied in real-time through co-contraction. Like the human arm it can therefore be ‘soft’ to absorb impacts, and to perform under uncertainty, but also stiffen up to be accurate. We take inspiration from human motor control to exploit these properties, and to try to make such arms practically useful. In particular, we are interested in performing fast and precise movements that are adapted to the task and environment. Our current architecture includes simple internal models that enable fast ballistic movements and collision detection. We would now like to explore the learning of such internal representations, and how they can be made sensitive to contextual clues.

 

Organized by

Verena Hafner