Upper Limb Pose–Configuration–Stiffness Transfer Framework for Serpentine Manipulator via Fusing Kinematic and Electromyography Signals Using Transformer Model

Upper limb skill transfer enables robots to acquire human-like operational capabilities by imitating human motion. However, transferring these skills to hyper-redundant serpentine manipulators presents considerable challenges due to the complex control of their flexible, continuous configurations. This paper proposes a novel upper limb skill transfer framework that facilitates simultaneous learning of pose, configuration and stiffness control. First, IMUs capture shoulder, elbow, and wrist poses of human operators. Joint positions are mapped through arm-length scaling. A bidirectional iterative inverse kinematics algorithm computes manipulator joint angles, achieving effective pose-configuration transfer. Second, muscle activation coefficients are integrated into the stiffness model of the manipulator, where the overall muscle contraction intensity governs directional stiffness variation through Null-space motion. Finally, an end-to-end skill transfer framework based on transformer is established. A multi-scale feature fusion strategy integrates low-dimensional IMU time-series data with spatiotemporal EMG matrices, effectively aligning and fusing bimodal representations to capture human operational skills. Experimental results demonstrate joint angle prediction with MSE = 4.6513°, R² = 0.9429, and PCC = 0.9769, significantly outperforming three baseline models.