When Does Physics Help? A Systematic Study of Physics-Guided Learning for Robotic Contact Dynamics

CHINMAYEE PRABHAKAR, Prathamesh Dinesh Joshi, Raj Dandekar, Rajat Dandekar, Sreedath Panat

Keywords: Physics-Informed Neural Networks (PINNs), Universal Differential Equations (UDEs), Scientific Machine Learning, Constrained Dynamical Systems, Zero-Shot Generalization, Parameter Identification

Abstract:

Robotic manipulation frequently involves contact with objects whose material properties are unknown, while force and state measurements are sparse, noisy, or unreliable. Learning accurate and physically valid contact dynamics under such conditions remains a core challenge.
Classical contact models rely on per-material parameter tuning and do not scale across heterogeneous objects, while purely data-driven models degrade under limited supervision and often violate physical constraints. Physics-informed neural networks (PINNs) and Universal Differential Equations (UDEs) provide promising alternatives, but it is unclear when and how physics-based inductive biases actually improve learning.
In this work, we conduct a systematic study of physics-guided learning for soft contact dynamics using a material-conditioned ODE formulation with explicit equality and inequality constraints. We compare data-only models, PINNs, and UDEs across controlled variations in data sparsity, noise, and temporal scope.
Our results show that while physics offers limited benefit in data-rich regimes, it becomes critical under sparse supervision: force errors reduce by up to 53%, friction-cone violations drop by up to 68%, zero-shot transfer to unseen materials improves by 14-29%, and physics-informed models match fully supervised performance within 5% using no material-specific training data.
These findings clarify the practical role of physics constraints as structured regularizers for reliable learning in real-world robotic contact scenarios.

Journal Opt In: No, I do not wish to participate

Submission Number: 140