Denoising Diffusion Planner: Learning Complex Paths from Low-Quality Demonstrations

TL;DR

It is shown that training a DDPM with synthetic and low-quality demonstrations is sufficient for generating nontrivial paths reaching arbitrary targets and avoiding obstacles, and different strategies for conditional sampling combining classifier-free and classifier-guided approaches are investigated.

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) are powerful generative deep learning models that have been very successful at image generation, and, very recently, in path planning and control. In this paper, we investigate how to leverage the generalization and conditional sampling capabilities of DDPMs to generate complex paths for a robotic end effector. We show that training a DDPM with synthetic and low-quality demonstrations is sufficient for generating nontrivial paths reaching arbitrary targets and avoiding obstacles. Additionally, we investigate different strategies for conditional sampling combining classifier-free and classifier-guided approaches. Eventually, we deploy the DDPM in a receding-horizon control scheme to enhance its planning capabilities. The Denoising Diffusion Planner is experimentally validated through various experiments on a Franka Emika Panda robot.

Figures (7)

• Figure 1: Overview of the DDP pipeline from low-quality and synthetic demonstration to real-world tasks.
• Figure 2: Forward process and reverse diffusion process for a path $\boldsymbol{\tau}$.
• Figure 3: Inpainting is used to set the start and goal poses of a generated path by fixing the start and goal poses (green square and yellow star, respectively) for all $k=K,K-1,...,0$.
• Figure 4: Block diagram of a closed-loop control configuration for a DDPM planner.
• Figure 5: Open-loop path generation with classifier-free guidance from the star to the diamond for various horizons $T$ and conditioned on various returns $c$. In (a), dense rewards are used and the returns are varied. In (b) sparse rewards are used and the returns are set to $c=0$. We generate thirty trajectories per combination of horizon and return and we highlight the mean trajectory with a blue line.