Tomorrow's Robots Learn Faster by Studying Physics, Not Just Data

Robotics is being reshaped by the rapid progress of artificial intelligence, accelerating the adoption of learning-based approaches. But while purely data-driven methods have achieved remarkable success in computer vision and natural language processing, robotics remains constrained by limited data, complex real-world interactions, and the need for reliable operation. That's the motivation behind physics-embedded robot learning: embedding physics priors into learning algorithms. By encoding underlying physical laws and constraints, physics priors can complement limited data with robotics-specific inductive biases, potentially improving generalization, interpretability, and sample efficiency.

The catch: the literature is fragmented across terminology, methodologies, and application domains, making this growing body of work difficult to assess. A new survey by Mattia Piccinini and 11 other authors reviews physics-embedded robot learning across a broad range of physics priors, robotics applications, and machine learning models, from single-layer perceptrons to generative foundation models.

The survey adopts a unified taxonomy classifying approaches by their physics embedding: physics-guided inputs, data, and representations; physics-encoded model architectures; and physics-informed training loss functions. Building on that, it reviews methods for robot dynamics learning, trajectory planning, prediction, control, and estimation, together with the corresponding open-source software ecosystem.

The authors' argument: physics priors provide a particularly relevant robotics-specific inductive bias, complementing rather than replacing data-driven learning, and paving the way toward more generalizable, data-efficient, and trustworthy robotic systems. They also identify key open challenges and outline promising future research directions.