Diffusion Policy for Collision Avoidance in a Two-Arm Robot Setup

Abstract

Diffusion models in robotics have shown great potential with a wide range of applicability. Especially their capability to model multi-modal data distribution has many benefits when learning a task such as collision-free trajectory optimisation with dynamic moving objects. Collision avoidance is a central problem in robotics where classical approaches still suffer from non-optimal solutions or high computational costs. In this work, we present four diffusion models, capable of predicting collision-free trajectories in a pick and place setup of two moving robot arms. The best model achieves a success rate of 88.1% in producing collision-free trajectories, while the worst one succeeds in 76.2% of the episodes. Furthermore, we analyse the models in terms of their accuracy in reaching the target pose, their capability of predicting smooth trajectories, and their success rate in generating collision-free trajectories.