Scaling PHAST to high-DOF robotics and continuum systems

Develop scalable methods to extend the PHAST (Port-Hamiltonian Architecture for Structured Temporal dynamics) framework to high-degree-of-freedom robotic systems and to continuum physical systems while preserving the port-Hamiltonian decomposition, passivity guarantees, and long-horizon forecasting stability under position-only (q-only) observability.

Background

PHAST is introduced as a unified port-Hamiltonian learning framework that models conservative and dissipative dynamics via explicit potential, mass, and damping components and uses structure-preserving integration. The paper evaluates PHAST primarily on q-only benchmarks with relatively low-dimensional systems across mechanical and non-mechanical domains.

In the Limitations, the authors note that performance in the q-only setting hinges on inferring momentum-like latents from short contexts and can degrade under severe sensor noise or partial observability. They explicitly state that scaling PHAST to high-DOF robotics and to continuum systems remains open, highlighting the need for algorithmic and computational advances to handle large state dimensions and more complex physics while maintaining physical guarantees.