Motion Composition and Interpolation Using Diffusion Models
Keywords: diffusion, composition, robotics
TL;DR: We show that the approximation used to compose diffusion models can be leveraged in robotics to generate novel motion by interpolating between specific modes in distributions.
Abstract: Humans have the ability to perform various combinations of skills without having to relearn the overall resulting skill every single time. For example, we prefer to learn easy motions and then combine them in flexible ways to perform complicated movements in dance. Enabling robots to combine or compose skills is essential for their deployment in unstructured environments where they will be required to adapt based on their surroundings. Without such composition robots would have to learn a separate policy for each task which can be combinatorially expensive. To this end, we propose a compositional approach to blend different robot skills using diffusion models. We compose configuration-space diffusion policies for novel motion generation resulting from the corresponding skill combinations. We show that the compositional framework can be utilized to interpolate between different skills leading to greater flexibility in motion. By utilizing interpolation along with composition, we can not only constrain the motion but also generate novel trajectories. We also propose a novel metric based on Maximum Mean Discrepancy and the Forward Kinematics kernel: MMD-FK to quantitatively evaluate the composed robot motion in the task-space while remaining agnostic to the space of policy composition.
Submission Number: 12
Loading