A Robust Open-source Tendon-driven Robot Arm for Learning Control of Dynamic Motions

Simon Guist, Jan Schneider, Hao Ma, Vincent Berenz, Julian Martus, Felix Grüninger, Michael Muehlebach, Jonathan Fiene, Bernhard Schölkopf, Dieter Büchler

Keywords: Reinforcement Learning, Dynamic Motions, Soft Robots, Pneumatic Artificial Muscles

Abstract:

Human athletic movements, such as those seen in table tennis or soccer, require a dynamic interplay of power, agility, and precision. They also involve quick reactions to unpredictable stimuli, while effectively managing impacts and maintaining safety during rapid interactions. Such complexities pose challenges for robots aiming to emulate or collaborate with humans in sports settings.
Most commercial robots are either precise yet fragile or safe yet underpowered. Tendon-driven robots provide a middle ground, lessening impact risks due to low inertia. Still, they often face precision issues due to unpredictable friction.
Our paper presents a newly designed 4-DoF tendon-driven robot arm, powered by pneumatic artificial muscles (PAMs), designed to minimize friction. This design achieves high force, low inertia, backdrivability, and superior precision compared to counterparts. While PAMs add complexities in control due to their nonlinear nature, reinforcement learning (RL) proves effective in handling them. Our robot's design also mitigates RL's collision risks during explorative training.
To foster further innovations, both the robot's hardware and software are made open-source.
The hardware predominantly employs readily available parts.
Our software offers an adaptable API in Python and C++ based on the o80 framework, interfacing with the robot's PLC over UDP. We will also open-source a huge proprioceptive data-set (25 days) including motion data at various speeds and forms.

Submission Number: 10