Session: SYMP S-6: Reservoir Computing & Nonlinear Dynamics

Paper Number: 139036

139036 - Correlating Mechanical Design to Physical Reservoir Computing Performance 

The ongoing advances in smart structure, functional materials, and soft robotics call for novel approaches to embody computation and intelligence directly in the mechanical domain. To this end, physical reservoir computing -- a non-traditional computing paradigm that treats high-dimensional and nonlinear mechanical systems as a recurrent neural network -- has shown many promising potentials. For example, our recent studies demonstrated that a simple origami structure can be a physical reservoir computer to accomplish state estimation and information perception. In this example, a Miura-ori- subject to base excitation at low frequency- behaves like a fixed neural network, and its vertice displacements are like the neuron signal. This way, training only involves a simple regression at the output layer. 

However, such training of physical reservoirs is still primarily an empirical effort, and the fundamental correlation between physical design, temporal dynamic response, and reservoir computing capacity remains elusive. This presentation will summarize our recent efforts to uncover such a physics-computing relationship.  

We selected three canonical systems as physical reservoirs: 

(1) A tapered spring: This is one of the most straightforward mechanical systems showing reservoir computing capability. It can be attached to a host, such as a wheeled autonomous vehicle, as a physically computing sensor to assess the ground conditions. 

(2) A 3D-printed TPU soft plate with embedded strain sensors: This plate exhibits highly nonlinear and large amplitude deformation under base excitation. Its interaction with the ambient environment, especially with fluid flows, has many applications for underwater robotics and energy harvesting. 

(3) Soft robotic modules with shape memory alloy actuators. These custom-made robotic modules have a plug-and-play architecture, where we can quickly swap the 3D-printed side panels and SMA spring actuators to adjust the overall stiffness. One can quickly assemble these modules into a high-degree-of-freedom soft manipulator. 

By working on these three systems, one can confidently apply this study's results and insights to different mechanical reservoirs. Specifically, we will use numerical simulation and extensive experiments to examine how changes in their physical design --- like geometry, material stiffness, and structure uniformity --- will influence a list of quantitative and measurable metrics based on their temporal responses. The metrics include nonlinearity, memory capacity, and dynamic symmetry. These spatial dynamics matrics can directly relate to physical computations and embodied intelligence performance. It is especially well-understood that there is a tradeoff between nonlinearity and memory in a physical reservoir, and we aim to use physical design to balance it. We hypothesize that, through these examinations, we can formulate mechanical design guidelines for different computing tasks, especially those relevant to soft robotic applications (such as state estimation for effective control). 

Presenting Author: Suyi Li Virginia Tech

Presenting Author Biography: Suyi Li is an associate professor of mechanical engineering at Virginia Tech. He has established a research program on Origami-inspired meta-structures and robotics. He is the recipient of the prestigious NSF CAREER award. He is also the recipient of the ASME Freudenstein Young Investigator Award, ASME Gary Anderson Early Career Award, and ASME C.D. Mote Jr Early Career Award.

Authors:

Jun Wang
Yogesh Phalak
Suyi Li