Session: SYMP S-5: Mechano-Intelligent Robots

Paper Number: 140025

140025 - Mechanical Metamaterial 'Brain' for Fully Analog Control of a Mobile Robot 

Physical Reservoir Computing (PRC) has been proposed as a promising route to harnessing the response of physical systems for distributed non-electronic computation. The PRC approach involves learning a linear read-out layer optimized to combine the physical nonlinearities that are suitable for the desired computation. Thus, a 'good' reservoir is a physical system that provides a large variety of nonlinear responses to be singled out by the read-out layer. In the context of mechanics, flexible mechanical metamaterials have emerged as a rich platform for nonlinear static and dynamic behaviors ranging from variable Poisson's ratio, buckling-induced pattern formation, and tunable stress-strain response, to soliton-like wave propagation. In this work, we draw inspiration from PRC principles to exploit the nonlinear response of a flexible metamaterial structure for sensing and robotic control.
First, we demonstrate that a flexible metamaterial structure — based on the rotating square mechanism — is capable of 'proprioception' when loaded on its boundary and its deformation is measured in just a few locations away from the surface. Specifically, the location at which the material is being loaded can be reconstructed from very limited strain information using a linear classifier. As such result can be achieved with a simple linear classifier, we then leverage this primitive sensing capability to realize a fully analog tactile control policy for directing the motion of a wheeled robot. Notably, this tactile control paradigm can be viewed as loosely inspired by the concept of thigmotaxis in biological locomotion i.e. the use of tactile information to direct the motion in living organisms.
By employing such a strategy, we design an autonomous mobile robot controlled exclusively by a flexible mechanical metamaterial without any digital electronics. The metamaterial is mounted onto a wheeled mobile base and is directly exposed to contact with the environment. Such deformable metamaterial 'skin' acts simultaneously as the sensory system that detects contact with obstacles and as the 'brain' that computes appropriate motor control voltages to free the robot from the obstacles and continue moving through the environment. In particular, as the robot runs into an obstacle, the metamaterial deforms nonlinearly due to the contact with the obstacle. A limited number of strain sensors measure the deformation and these strain measurements are linearly superimposed with experimentally trained weights through an analog circuit to produce the adequate motor control command. No digital computation is required. The 'intelligence' and sensing capability of the robot is directly outsourced to the nonlinearly deforming mechanical metamaterial.
In light of these results, we envision that flexible mechanical metamaterials can be a promising platform for the physical embodiment of complex computation in robotic applications.

Presenting Author: Giovanni Bordiga Harvard University

Presenting Author Biography: Giovanni Bordiga is a Postdoctoral Fellow at Harvard School of Engineering and Applied Sciences. He received a PhD in Solid and Structural Mechanics from the University of Trento (Italy). His research focuses on material instabilities in structured media, homogenization, linear and nonlinear waves, metamaterials, mechanical intelligence, and simulation-driven inverse design.

Authors:

Cyrill Bösch
Giovanni Bordiga
Connor Mccann
Sina Jafarzadeh
Jackson Wilt
Michelle Yuen
Yichu Jin
Eder Medina
Andreas Fichtner
Katia Bertoldi