Integration of Force-Torque Information into Reinforcement Learning for Accurate Position Estimation Using the TD3 Algorithm

Robotic manipulation tasks often require precise position estimation, a crucial aspect for successful task completion. In this study, we propose a novel approach to enhance such position estimation by integrating force-torque information into the reinforcement learning (RL) framework. Unlike conventional methods that heavily rely on visual data, our approach focuses exclusively on leveraging force-torque data. Specifically, we utilize the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, known for its effectiveness in RL tasks. By incorporating force-torque feedback, our method offers a promising alternative for achieving accurate position estimation in robotic tasks, especially in scenarios where visual data may be limited or unreliable. To evaluate the effectiveness of the RL-based positioning strategy, the experiment is conducted using a simulation environment. The Mujoco physics engine is employed to simulate the peg-in-hole assembly task and assess the performance of the learned control policy.