Back to Blog
Eduardo Okabe, Victor Paiva, Luis Silva-Teixeira, Jaime Izuka
J. Comput. Nonlinear Dynam. Oct 2023, 18(10): 104501
The industry and the scientific community have shown interest in SCARA (Selective Compliance Assembly Robot Arm) robots due to their high accuracy. In this paper, a two-link SCARA has its end-effector pulled by three cables that generate a triangular-shaped workspace. Moving the end-effector in this region is a relatively straightforward task, but placing the end-effector outside it requires a nonlinear dynamic model and a state-of-the-art controller. To address this problem in a simpler, more efficient and innovative manner, the equations of motion are derived and three reinforcement learning algorithms are employed: Proximal Policy Optimization (the same used by the chatbot ChatGPT), Soft Actor-Critic and Twin Delayed Deep Deterministic Policy Gradient. Three targets outside the triangular workspace are considered and the trained networks have their results compared in terms of displacement error, velocity and standard deviation. The Twin Delayed Deep Deterministic Policy Gradient provides creative trajectories, the Soft Actor-Critic presents better solutions for two out of three targets, while the Proximal Policy Algorithm appears to be the most consistent considering all targets under analysis.