Evolving Modular Soft Robots without Explicit Inter-Module Communication using Local Self-Attention

Abstract

Self-attention visualization GeneralizationFigure 1: Overview of the proposed approach. We use the same neural controller (left picture) inside each voxel, with shared parameters. The middle picture is a biped with the attention matrices of the different voxels. Each controller uses self-attention to compute importance scores (𝑨) among the inputs sensed by its voxel. We also find evolved controllers to generalize to unseen morphologies (right picture; color represents the ratio between the voxel current area and its rest area: red stands for contraction, green for expansion, yellow for no change).

Reconfigurable, Multi-Material, Voxel-Based Soft Robots

“…The presented reconfigurable VSR design and assembly method can be used to perform future research on evolved VSRs while reducing material costs, manufacture labor and waste production. More specifically, since research on simulated VSRs is moving towards closed-loop control (based on artificial neural networks) [29]- [31], future work will focus on integrating self-healing sensors [32] inside the voxels to perform closed-loop control of real reconfigurable, multimaterial VSRs.…”