Embedding Physics Priors in Robot Learning: A Survey

This survey reviews physics-embedded robot learning by classifying existing approaches into a unified taxonomy: physics-guided inputs, data, and representations; physics-encoded model architectures; and physics-informed training loss functions. Covering applications from robot dynamics learning to trajectory planning, control, and estimation, the survey highlights how embedding physical laws complements data-driven methods with robotics-specific inductive biases to enhance generalization, interpretability, and sample efficiency.

We adopt a unified taxonomy that classifies existing approaches according to their physics embedding: physics-guided inputs, data, and representations; physics-encoded model architectures; and physics-informed training loss functions.