Creation of Novel Soft Robot Designs using Generative AI

TL;DR

The paper tackles the challenge of designing soft robot actuators by leveraging generative diffusion models trained on a dataset of over 70 text–shape pairs to produce 3D soft actuator geometries. Building on SDFusion, it employs a VQ-VAE latent space and a text-conditioned latent diffusion framework, enhanced by transfer learning and data augmentation to improve fidelity and surface quality. They introduce and apply multiple evaluation metrics (FID, 3D Fréchet Distance, surface smoothness, and volumetric IoU) and demonstrate that the approach can generate novel, shape-rich designs, albeit with limitations in resolution and physics integration. The work suggests that scaling data and hardware, along with integrating physics via simulation or reinforcement learning, could enable rapid, co-optimized morphology–control design for soft robotics.

Abstract

Soft robotics has emerged as a promising field with the potential to revolutionize industries such as healthcare and manufacturing. However, designing effective soft robots presents challenges, particularly in managing the complex interplay of material properties, structural design, and control strategies. Traditional design methods are often time-consuming and may not yield optimal designs. In this paper, we explore the use of generative AI to create 3D models of soft actuators. We create a dataset of over 70 text-shape pairings of soft pneumatic robot actuator designs, and adapt a latent diffusion model (SDFusion) to learn the data distribution and generate novel designs from it. By employing transfer learning and data augmentation techniques, we significantly improve the performance of the diffusion model. These findings highlight the potential of generative AI in designing complex soft robotic systems, paving the way for future advancements in the field.

Figures (8)

• Figure 1: The network comparison between traditional autoencoder, variational autoencoder kingma2013auto and vector quantised variational autoencoder van2017neural. The picture is drawn by us by referred from the original papers.
• Figure 2: Diffusion process for a voxel represented shape. the forward process is to constantly add Gaussian noise, and the reverse process is learned by training a denoiser. In this report, the true diffusion process happens in latent space instead of voxel space, this is only for illustration of diffusion process. The original idea is from ho2020denoising.
• Figure 3: Structure of latent diffusion model rombach2022high.
• Figure 4: Soft pneumatic robot actuator designs used for dataset.
• Figure 5: Text-Shape Pairings of Soft Actuators.