TD3 Integrated Fuzzy-finite Variable Admittance Control of Posture Estimation and Adjustment for Robotic Precise Peg-in-hole

In unstructured environment, the robot faces the Precise Peg-in-Hole (PPiH) assembly as a non-cooperative issue. The posture uncertainty of the peg presents challenges in searching and inserting the hole. The purpose of the research is to eliminate the posture deviation between the peg and the hole with the force feedback. In this paper, the posture adjustment is divided into rough and fine processes. Firstly, for the rough adjustment, the force-angle samples from the end-effector are trained using a Multi-Layer Perceptron (MLP) model under peg-hole non-contact. The robot is guided by the MLP and adjusts the peg to roughly compensate for the posture deviation. Then, the robot brings the peg toward and contacts the hole. Secondly, for the fine adjustment, a Fuzzy-Finite Variable Admittance Control (FFVAC) model is established to estimate and adjust posture for different peg-hole contact states accurately. By integrating force information with fuzzy logic, the fuzzy inference system with fuzzy rules is developed based on the peg-hole contact states. According to the contact states, the twin delayed deep deterministic policy gradient (TD3) model finds the optimal admittance control parameters to achieve the surface close fitting of the peg and hole. Finally, comprehensive experiments are conducted under the unknown initial posture of the peg. The results are analysed by comparing with the stated of the arts of the posture adjustment methods regarding adjustment accuracy and operation time. The proposed method quickly reduces the posture deviation angle less than 0.2°, facilitates the following hole search and insertion works.