Physics-Informed Learning Framework for Port-Hamiltonian Systems
A novel framework for physics-informed learning simultaneously develops port-Hamiltonian (pH) system models and optimal energy-balancing passivity-based controllers (EB-PBC) using trajectory data. This method involves iteratively improving the system model based on data gathered under the existing control policy and subsequently re-optimizing the controller according to the revised model. Both elements are represented by neural networks that incorporate pH dynamics and the EB-PBC framework, providing clarity regarding energy interactions. The resulting controller ensures that the closed-loop system remains inherently passive and demonstrably stable, leveraging passive plant dynamics without negating natural potential. Additionally, a dissipation regularization technique promotes strict energy decay during training, bolstering resilience against sim-to-real discrepancies.
Key facts
- Framework co-learns pH system models and EB-PBC controllers from trajectory data.
- Alternating optimization with policy-aware data collection is used.
- Neural networks parameterize both model and controller, embedding pH dynamics and EB-PBC structure.
- Learned controller ensures closed-loop passivity and provable stability.
- Controller exploits passive plant dynamics without canceling natural potential.
- Dissipation regularization enforces strict energy decay during training.
- Approach enhances robustness to sim-to-real gaps.
- Published on arXiv with ID 2604.26172.
Entities
—