A Soft Robotic System Automatically Learns Precise Agile Motions Without Model Information

TL;DR

The paper tackles the challenge of precisely controlling Pneumatic Soft Actuators (PSAs) despite nonlinearities and hysteresis. It introduces Automatic Neural ODE Control (ANODEC), a two-stage data-driven approach that first learns a neural ODE model of the plant from input-output data and then synthesizes a neural ODE feedback controller, both optimized via RK4 integration and Adam optimization. On a real PSA with hysteresis, ANODEC achieves agile, non-repetitive reference tracking using only 30 s of interaction data and outperforms a manually tuned PID baseline across multiple unseen reference signals and a second, gravity-affected setup. This demonstrates data-efficient, model-free automatic control design for practical soft robotics, enabling robust agile motions with minimal experimental interaction time. Future work includes extending to multi-DOF PSAs and automating hyperparameter tuning and data acquisition.

Abstract

Many application domains, e.g., in medicine and manufacturing, can greatly benefit from pneumatic Soft Robots (SRs). However, the accurate control of SRs has remained a significant challenge to date, mainly due to their nonlinear dynamics and viscoelastic material properties. Conventional control design methods often rely on either complex system modeling or time-intensive manual tuning, both of which require significant amounts of human expertise and thus limit their practicality. In recent works, the data-driven method, Automatic Neural ODE Control (ANODEC) has been successfully used to -- fully automatically and utilizing only input-output data -- design controllers for various nonlinear systems in silico, and without requiring prior model knowledge or extensive manual tuning. In this work, we successfully apply ANODEC to automatically learn to perform agile, non-repetitive reference tracking motion tasks in a real-world SR and within a finite time horizon. To the best of the authors' knowledge, ANODEC achieves, for the first time, performant control of a SR with hysteresis effects from only 30 seconds of input-output data and without any prior model knowledge. We show that for multiple, qualitatively different and even out-of-training-distribution reference signals, a single feedback controller designed by ANODEC outperforms a manually tuned PID baseline consistently. Overall, this contribution not only further strengthens the validity of ANODEC, but it marks an important step towards more practical, easy-to-use SRs that can automatically learn to perform agile motions from minimal experimental interaction time.

Figures (6)

• Figure 1: Comparison between conventional controller design, reliant on human expertise, and an automatic, data-driven control structure for Pneumatic Soft Actuators (PSAs). Both controller design paradigms observe the PSA's control signal and resulting measured motion, and subsequently design position feedback controllers that enable the measured motion of the PSA to track the desired motion. In this work we show that the data-driven method, Automatic Neural ODE Control (ANODEC), enables automatic control design for a PSA.
• Figure 2: The PSA including two PAMs is controlled using two pressure controllers, which implement a SISO control strategy through the mean and difference pressure approach. As a result, the single control input $u$ controls both desired difference pressures $p_\text{d,1/2}$, and the single system output is the hinge joint angle $\varphi$.
• Figure 3: Training data for ANODEC consists of five input-output pairs, each of a five seconds length, that are gathered from the experimental PSA. One additional input-output pair (dashed line) is collected and used as validation data and to prevent model overfitting. The feasible input and output intervals of the experimental PSA are $u(t) \in [-6, 6] \, \qty{}{\bar{}}$ and $\varphi(t) \in [-1, 1] \, \qty{}{\radian}$, respectively.
• Figure 4: Two experimental setups of a pneumatic arm with a single DOF. In both setups, two PAMs (pneumatic artificial muscles, black tubes) are used as an antagonistic pair to control the arm's forces and position. The upper setup shows the simplest configuration without the influence of gravity and external load. The lower setup is loaded with an external weight of 0.6kg, with a lever arm of 0.25m oriented against gravity. Both arms are fixed to the ground to prevent undesired movements.
• Figure 5: Performance comparison of the automatic ANODEC and a manually tuned PID controller baseline in Setup 1 for one exemplary reference signals drawn from the three reference signal distributions: Step, double step, and smooth. Even on reference signals that were not present during training (double step, smooth), ANODEC is able to consistently outperform the PID baseline and achieves a lower RMSE tracking error while requiring less experimental interaction time. Video (https://youtu.be/7HkXKy0WuRw) that showcases these trials.