Abstract

Classical methods for solving Inverse Kinematics (IK) problems result in non-unique solutions, especially for higher degrees of freedom robotic arm. Also, these methods are computationally intensive and non-trivial. To circumvent this, Deep Reinforcement Learning (DRL) methods have shown potential to address these challenges in the recent literature. Thus, in this work, we undertake a DRL approach for a representative 7-DoF Panda Franka Emika robot. Moreover, the literature misses out on a comprehensive study of multiple DRL methods for examining the key performance parameters of the IK problem for a 7-DoF cobot. The performance of three popular DRL algorithms- Proximal Policy Optimization (PPO), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Deep Deterministic Policy Gradient (DDPG) is performed for this purpose. We evaluate their performance using key indicators such as accuracy in terms of average position error, success rate, and convergence time. Moreover, all these methods are tested against the parametric uncertainties applied in multiple links for the IK problem, inferring its practicality and robustness in real-world scenarios. In due course, the training framework is also proposed through which a particular DRL algorithm is trained with reward logs that is used to evaluate the training stability and learning progress. Finally, for each DRL algorithm, we remark a few quantitative metrics that suggest their selection guidelines for a given scenarios in specific IK problem.

Key words: Inverse Kinematics; Deep Reinforcement Learning; PPO; TD3; DDPG; Robotic Manipulation.