Diffusion Models Improve Robot Control in Challenging Friction Environments

Diffusion models, known for their generative capabilities, are now being investigated for their potential in planning and control systems. This research specifically examines "Action Diffusion," a formulation that uses diffusion models to generate action sequences for open-loop control. The study focuses on a challenging scenario: controlling a point-mass system subject to dry friction and stiction, where motion only begins after a static friction threshold is overcome. The proposed approach utilizes a compact conditional 1D U-Net to create bounded control sequences, conditioned on the initial and target states. This allows the model to learn and recombine structured control primitives from its training data. Compared to methods like uniform random shooting, dataset-prior random shooting, and the Cross-Entropy Method (CEM), Action Diffusion demonstrates superior performance. It significantly reduces terminal error and the number of "stuck" steps, particularly when data samples are limited. These findings suggest that conditional diffusion models are highly effective for generating precise, temporally coherent control sequences, especially in environments with complex physical constraints like stiction.