Information processing via physical soft body

Abstract

Soft machines have recently gained prominence due to their inherent softness and the resulting safety and resilience in applications. However, these machines also have disadvantages, as they respond with complex body dynamics when stimulated. These dynamics exhibit a variety of properties, including nonlinearity, memory, and potentially infinitely many degrees of freedom, which are often difficult to control. Here, we demonstrate that these seemingly undesirable properties can in fact be assets that can be exploited for real-time computation. Using body dynamics generated from a soft silicone arm, we show that they can be employed to emulate desired nonlinear dynamical systems. First, by using benchmark tasks, we demonstrate that the nonlinearity and memory within the body dynamics can increase the computational performance. Second, we characterize our system's computational capability by comparing its task performance with a standard machine learning technique and identify its range of validity and limitation. Our results suggest that soft bodies are not only impressive in their deformability and flexibility but can also be potentially used as computational resources on top and for free.

Figure 1. Platform setup for a soft silicone arm and schematics showing the information processing scheme using the arm.

(a) Schematics showing a scheme using the soft silicone arm as a part of a computational device. The arm embeds 10 bend sensors and is immersed underwater. Inputs are motor commands that generate arm motions, and the embedded bend sensors reflect the arm posture for each timestep. Corresponding system outputs are generated by the weighted sum of the sensory values. (b) Picture showing the soft silicone arm used in this study. (c) Schematics expressing an analogy between a conventional reservoir computing system and our system. In a conventional reservoir system, randomly coupled abstract computational units are used for the reservoir, whereas our system exploits a physical reservoir whose units are sensors that are coupled through a soft silicone material. Our question here is whether our physical reservoir can perform tasks of nonlinear dynamical systems emulations, which are often targeted with conventional reservoir computing systems and are useful in the context of control.

Figure 2. A typical example of the task performance in terms of time series when T = 400 in the evaluation phase.

From the upper to the lower plots, the time series of the motor command, the corresponding sensory values (odd-numbered sensors and even-numbered sensors), and the outputs for NARMA2, NARMA5, NARMA10, NARMA15, and NARMA20 are depicted. For each plot of the output, the time series for the system output as well as the target output and the output for the LR model is overlaid for comparison. Note that the output of the LR model, especially in NARMA5, NARMA10, NARMA15, and NARMA20, is not a constant but a scaled version of the input with an offset (see the inset that scales up the output for the LR model from timestep 6300 to 6400 in each plot).

Figure 3. Comparisons among the average NMSE_system, NMSE_LR, , and for all NARMA tasks for each setting of T (a) and a diagram summarizing the significant differences between NMSE_system and (b).

In (a), error bars represent standard deviations. For each task, the average NMSE_system is significantly lower than the average NMSE_LR (seemingly overlapping plots in, e.g., the NARMA10 task with T = 100 and 150, the NARMA15 task with T = 200, and the NARMA20 task with T = 100 and 250, are due to the scaling of the figures), while the average is significantly lower than the average NMSE_system for each setting of T (Supplementary Table S1). In (b), among NMSE_system and , the significantly lower one with p < 0.05 is depicted for each experimental condition. Note that “n.s.” represents “not significant.” All the information, including the average NMSEs as well as the results for significant tests in each experimental condition, is given in Supplementary Table S1.