Generalizable and Fast Surrogates: Model Predictive Control of Articulated Soft Robots Using Physics-Informed Neural Networks
Tim-Lukas Habich, Aran Mohammad, Simon F. G. Ehlers, Martin Bensch, Thomas Seel, Moritz Schappler
AI summary
Problem
First-principles models are computationally too slow for real-time control, while black-box learned models fail to generalize to unseen system changes and require extensive real-world data.
Approach
The authors train physics-informed neural networks using minimal real-world data from a single operating domain, embedding governing physical laws as constraints to force the model to extrapolate accurately to new dynamics.
Key results
- PINN prediction speed exceeds accurate first-principles models by up to 467x with only slightly reduced accuracy.
- PINN demonstrates high generalizability to unseen dynamics (changed payload and orientation) compared to recurrent neural networks.
- Enables real-time nonlinear model predictive control (NMPC) at 47 Hz with accurate position tracking in dynamic experiments.
- Achieves robust control performance without retraining the model or retuning the controller across different system domains.
Why it matters
Provides a scalable, data-efficient modeling framework that bridges the accuracy-speed-generalization gap, enabling real-time MPC for complex soft robots in dynamic real-world environments.
Abstract
Soft robots can revolutionize several applications with high demands on dexterity and safety. When operating these systems, real-time estimation and control require fast and accu- rate models. However, prediction with first-principles (FP) models is slow, and learned black-box models have poor generalizability. Physics-informed machine learning offers excellent advantages here, but it is currently limited to simple, often simulated systems without considering changes after training. We propose physics-informed neural networks (PINNs) for articulated soft robots (ASRs) with a focus on data efficiency. The amount of expensive real-world training data is reduced to a minimum — one dataset in one system domain. Two hours of data in different domains are used for a comparison against two gold- standard approaches: In contrast to a recurrent neural network, the PINN provides a high generalizability. The prediction speed of an accurate FP model is exceeded with the PINN by up to a factor of 467 at slightly reduced accuracy. This enables nonlinear model predictive control (MPC) of a pneumatic ASR. Accurate position tracking with the MPC running at 47 Hz is achieved in six dynamic experiments.