Embedding Physics Priors in Robot Learning: A Survey

This survey examines the intersection of physics priors and robot learning, addressing how domain-specific physical knowledge can be embedded into learning frameworks to overcome data scarcity and ensure reliable real-world operation. The review covers approaches including physics-informed neural networks, constraint-based reinforcement learning, and hybrid model-based methods that incorporate multibody dynamics, contact mechanics, and rigid/flexible body coupling into the learning pipeline.