This paper presents a Diffusion-Based Impedance Learning framework that combines learning methods that excel at motion generation in the information domain with impedance control that shapes physical interactions in the energy domain. A Transformer-based diffusion model is used to reconstruct a simulated Zero-Force Trajectory (sZFT), and a SLERP-based quaternion noise scheduler is introduced for rotation to ensure geometric consistency. The reconstructed sZFT is then fed into an energy-based estimator that updates the stiffness and damping parameters. Data were collected for park play scenarios and robot-assisted therapy tasks, achieving accurate position and rotational accuracies even with a small sample size. The small model size enables real-time torque control and autonomous stiffness adaptation, and successful results are achieved in various peg insertion tasks.
