SAC-Optimized Fuzzy Variable Admittance Control for Lead-Through Teaching of Collaborative Robots

In collaborative robot lead-through teaching, fixed admittance parameters impose an inherent trade-off between operational ease and motion stability. This paper proposes a SAC-optimized fuzzy variable admittance control method (SAC-FAC). A fuzzy variable admittance controller (FAC) quantifies the operator's motion and turning intents using interaction-force and end-effector motion information, and modulates the damping coefficient online via interpretable fuzzy rules. Soft Actor-Critic (SAC) searches offline for a well-balanced membership-function configuration on an episode basis in simulation, and the optimized configuration is then fixed for deployment. A saturation mechanism in the reward function suppresses degeneration of the damping configuration toward its physical lower bound. To counter parameter degradation under high-disturbance training, potential-based reward shaping and performance-gated curriculum learning are jointly introduced to promote stable convergence. Ablation studies and comparisons with four alternative optimizers verify the training design and support the suitability of SAC in this framework. Experiments on a UR10 collaborative robot platform with three trajectory types show that, relative to the hand-tuned FAC, SAC-FAC reduces the mean trajectory tracking error, work per unit path, and root-mean-square interaction force by 19.5%, 11.6%, and 6.8%, respectively, with more evident advantages on the compound and 3D ramp trajectories while preserving the interpretability of the fuzzy rule structure.