Diffusion-Based Impedance Learning for Contact-Rich Manipulation Tasks

This framework combines diffusion-based generative models with physics-informed control to enable adaptive behavior in contact-rich manipulation tasks. The system is trained on human demonstrations captured through Apple Vision Pro teleoperation, where subjects guide the robot through physical environments under varying interaction conditions. A Transformer-based conditional diffusion model denoises both translational and rotational data, using quaternions to represent and reconstruct orientation. During deployment, the robot adapts its impedance in real time based on observed motion and force patterns. This approach enables safe, compliant, and task-specific interaction, with applications ranging from physical therapy to contact-rich manipulation in industrial and service robotics.

Teleoperated Data Collection for Model Training

1. Human-Guided Upper Limb Therapy Demonstration via Apple Vision Pro
Teleoperation is used to record patient movements with varying levels of support, enabling the model to learn the dynamic relationship between force and motion.

2. Parkour Trajectory Demonstration via Teleoperation
Teleoperation is used to guide the robot through physical obstacles, enabling the model to learn how interaction forces vary with pose and context.


Model Deployment in Real-World Contact Tasks

1. Real-Time Adaptive Support in Upper Limb Therapy Deployment
The robot autonomously assists with upper-limb rehabilitation by adjusting support in real time based on patient engagement.

2. Contact-Rich Robot Deployment in a Parkour Environment
The robot performs a parkour task, adapting its impedance in real time to manage unknown contact-rich interactions.


N. Geiger, C. Piazza, T. Asfour, N. Hogan, J. Lachner, “Physics-Grounded Diffusion for Therapist-Like Robot Support in Upper-Limb Stroke Rehabilitation,” In preparation, 2025.

N. Geiger, T. Asfour, N. Hogan, J. Lachner, “Diffusion-Based Impedance Learning for Contact-Rich Manipulation Tasks,” In preparation, 2025.

Acknowledgements

This work was developed together with Noah Geiger and Prof. Tamim Asfour.