High-Performance Series Elastic Actuator Design and Reinforcement Learning Control Based on the DDPG Algorithm

The Series Elastic Actuator (SEA) has garnered significant attention in the field of robotic actuation due to its inherent passive compliance, which enhances safety and energy efficiency in human-robot interaction. However, its widespread deployment remains constrained by limitations in torque density and the difficulty of achieving high-precision control in the presence of model uncertainties and nonlinear dynamics. To address this, this paper presents a novel high-torque-density SEA joint design and establishes a model-free reinforcement learning control framework based on the Deep Deterministic Policy Gradient (DDPG) algorithm. We developed and trained a DDPG-based control strategy in a simulation environment, where the adoption of compound dynamic reference trajectories ensures the robustness and broad adaptability of the trained agents. Experimental results show that the DDPG controller outperforms traditional methods-including PD with disturbance observer (PD+DOB) and linear quadratic regulators (LQR)-in various trajectory tracking tasks, demonstrating its effectiveness and superiority for high-performance, robust, and model-free control of SEA systems.