Diffusion-Based Impedance Learning Enhances Contact-Rich Robot Manipulation

The recently introduced framework, Diffusion-Based Impedance Learning, merges generative modeling with energy-consistent impedance control to enhance robot capabilities during contact-intensive interactions. This innovative approach, outlined in a paper on arXiv (2509.19696), tackles the shortcomings of learning-based techniques in tasks involving contact and the necessity for specific tuning in impedance control. Utilizing a Transformer-based Diffusion Model, the framework employs cross-attention on external wrenches to recreate simulated Zero-Force Trajectories (sZFTs), which embody equilibrium behavior during contact. To maintain geometric consistency for rotations, a SLERP-based quaternion noise scheduler is implemented. The reconstructed sZFT informs an energy-based estimator that modulates impedance in real-time through directional stiffness and damping adjustments. Training was conducted using demonstrations from parkour and robot-assisted therapy via a teleoperation setup. This paper, categorized as a replace-cross type, was published on arXiv on September 25, 2025, and can be accessed at https://arxiv.org/abs/2509.19696.