Quantification of Joint Redundancy considering Dynamic Feasibility Using Deep Reinforcement Learning

The robotic joint redundancy for executing a task and the optimal usage of robotic joints given the redundant degrees of freedom are crucial for the performance of a robot. It is therefore of interest to quantify the joint redundancy to better understand the robotic dexterity considering the dynamic feasibility. To this end, model-based approaches have been among the most commonly used methods to quantify the joint redundancy of simple robots analytically. However, this classical approach fails when applied to non-conventional complex robots. In this study, we propose a new method based on a deep reinforcement learning-derived metric, the synergy exploration area (SEA) metric, for the quantification of redundancy with a given dynamic environment. We conducted various experiments with different robotic structures for different tasks, ranging from simple robotic arm manipulation to more complex robotic locomotion. The experimental results show that the SEA metric can effectively quantify the relative joint redundancy over different robotic structures with varying degrees of freedom under unknown dynamic situations.