Re-purposing a modular origami manipulator into an adaptive physical computer for machine learning and robotic perception

TL;DR

The paper addresses how mechanical design governs computing performance in embodied physical computers by repurposing a modular origami manipulator as an adaptive physical reservoir. It uses a fixed readout layer with trainable weights to map high-dimensional body dynamics, captured from 40 markers, to task outputs via $O(t)=w_0+\sum w_i s_i(t)$, and evaluates performance on NARMA time-series emulation, payload weight estimation, and SMA-driven robotic multitasking. Key contributions include introducing Peak Similarity Index (PSI) and spatial correlation as design-guiding metrics, demonstrating configuration- and input-dependent computing capacity, and showing practical information extraction and control tasks enabled by the physical kernel, including payload perception and exteroception via SMA actuation. The findings illustrate how adaptive, reconfigurable mechanical structures can enable embodied intelligence, providing a framework for future soft robotics and bio-inspired adaptive materials to compute and interact with digital counterparts in the mechanical domain.

Abstract

Physical computing has emerged as a powerful tool for performing intelligent tasks directly in the mechanical domain of functional materials and robots, reducing our reliance on the more traditional COMS computers. However, no systematic study explains how mechanical design can influence physical computing performance. This study sheds insights into this question by repurposing an origami-inspired modular robotic manipulator into an adaptive physical reservoir and systematically evaluating its computing capacity with different physical configurations, input setups, and computing tasks. By challenging this adaptive reservoir computer to complete the classical NARMA benchmark tasks, this study shows that its time series emulation performance directly correlates to the Peak Similarity Index (PSI), which quantifies the frequency spectrum correlation between the target output and reservoir dynamics. The adaptive reservoir also demonstrates perception capabilities, accurately extracting its payload weight and orientation information from the intrinsic dynamics. Importantly, such information extraction capability can be measured by the spatial correlation between nodal dynamics within the reservoir body. Finally, by integrating shape memory alloy (SMA) actuation, this study demonstrates how to exploit such computing power embodied in the physical body for practical, robotic operations. This study provides a strategic framework for harvesting computing power from soft robots and functional materials, demonstrating how design parameters and input selection can be configured based on computing task requirements. Extending this framework to bio-inspired adaptive materials, prosthetics, and self-adaptive soft robotic systems could enable next-generation embodied intelligence, where the physical structure can compute and interact with their digital counterparts.

Figures (21)

• Figure 1: The overview of this study. (a) A reservoir computer features a fixed neural network and a simple output layer of weighted linear summation. (b) In this study, we employ a reconfigurable robotic arm consisting of origami-inspired, meta-stable modules as the physical platform. (c) By applying the reservoir computing framework to this robotic arm, we creat a physical computer with an adaptive kernel, which can be dynamically excited with base excitation or embedded shape memory alloy (SMA) coil actuators. The dynamics of this adaptive kernel are represented by the displacements of the markers attached throughout its body, which function as the reservoir state vectors $s_i(t)$. The reservoir's computing output is simply a weighted linear summation of these nodal displacements $\mathbf{O}(t)=w_0+\sum w_i s_i(t)$, where the constant readout weights $w_i$ will be trained according to the tasks at hand. (d) In the big picture, this study reveals that the spectral and spatial correlations between the reservoir states and the targeted outcome is an informative hidden layer between the mechanical configuration of a physical computer and its computing performance.
• Figure 2: Fabrication and configuration of the origami-inspired meta-stable modules. (a) To fabricate a bistable origami panel, we first 3D-printed it using flexible TPU and stiffer Nylon filaments and heat-treat it to reset its stress-free folding configurations. After that, we fit the origami panel with a coil spring or an SMA actuator coil, giving it the desired bistability featuring soft [0] and stiff [1] states. (b) We assembled three such panels and two 3D-printed end plates to complete the meta-stable module as the fundamental element in our adaptive physical computing kernel. Here, we focus on the most compliant [000] and stiff [111] states. (c) Tensile and compression testing reveal that the longitudinal stiffness ratio between the [111] and [000] states is $4.00\pm0.15$. The solid lines are averaged test results from 10 loading cycles, and the shaded bands are the corresponding standard deviation. (d) Close-up view of the six other metastable states from the same module.
• Figure 3: A summary of the computing tasks in this study, ranging from the more fundamental task (I) to the more practical multi-tasking (III). We investigated these three tasks using different physical configurations ($\mathbf{C}_1$ to $\mathbf{C}_8$) by adapting the number and stable states of origami modules. In the time series emulation task (I), the input to the adaptive kernel is a base excitation $\mathbf{I}^{(\text{I})}(t)$ consisting of three harmonic signals; and the targeted output is a time series defined by nonlinear NARMA equations. In the information extraction task (II), we attached masses to the adaptive kernel's free end. The input is a simple harmonic base excitation with different frequencies, and the targeted output is the weight of these end mass. In the final multi-tasking (III), we replaced the embedded passive coil springs with active SMA coils and input Pulse Width Modulated (PWM) current to these SMA to excite (swing) the kernel. The output targets are to reconstruct the PWM input commands, classify the payload by estimating its weight, and determine the orientation of a payload.
• Figure 4: The adaptive kernel's computing performance using the NARMA benchmark. The reservoir dynamics differs, with different physical configurations $\mathbf{C}_m$ and base excitation magnitude $\mathbf{A}_n$. (a) The nodal displacements of $\mathbf{C}_1$ and $\mathbf{C}_5$ under different input magnitudes. (b) From the left to the right columns: The temporal outcome, spectral comparison, and NMSE/PSI summary of for NARMA2, NARMA5, NARMA10, NARMA15, and NARMA20 tasks. (c) The adaptive kernel's optimal configuration corresponding to different NARMA tasks and input magnitudes. The values on this color map is the percentage reduction of NMSE error compared to the $\mathbf{C}_5$ configuration performance at that input magnitude.
• Figure 5: Extracting payload weight from body dynamics, using different kernel configurations and input conditions. (a) The adaptive kernel's body dynamics are represented by nodal displacements. (b) The correlation matrix between nodal displacements provides a quantitative measure of the "richness" of a physical body's dynamic characteristics. (c) The adaptive kernel estimates its payload mass using different amounts of readout training data. If one uses the data from all five payloads for readout training, the kernel can output very accurate weight estimation. On the other hand, if one only uses the data from 1 payload, the kernel will fail the estimation task. (d) The reservoir's prediction using readout weights trained using data from two payloads. These results are accurate only when the spatial correlation between nodal displacements drops. (e) Averaged spatial correlation between different physical and input settings. (For visual clarity, only the lower left sub-plot's axes are labeled in b-d, and the same scale applies to all other sub-plots) (f) The overall estimation error in all testing cases with different readout training setups.